meta analysis of enhanced efficiency fertilizers in corn...
TRANSCRIPT
Page 1 of 66
Meta-analysis of Enhanced Efficiency Fertilizers in Corn Systems in the Midwest
Rachel Cook1, Amy Nail2, April Vigardt1, Andrew Trlica3, Brooke Hagarty1, Travis Williams1, Jeff Wolt4
1Southern Illinois University-Carbondale, 2Honestat, LLC, 3Boston University, 4Iowa State University
Executive Summary Enhanced efficiency fertilizers such as nitrification inhibitors, urease inhibitors, and polymer
coated fertilizers may provide environmental and agronomic benefits in nitrogen fertilization of
corn production systems. Their effectiveness, however, can be influenced by agricultural
management practices and can vary by location due to climate and soil characteristics. In this
study we performed a systematic review of published data from the Midwest to address the
effectiveness of enhanced efficiency fertilizers on corn yield, nitrous oxide emissions, and nitrate
leaching. We identified 46 studies comprising 1248 observations which were amenable to
meta-analyses describing corn yield response for the years 1994-2014. We modeled corn yield
data by primary fertilizer sources (anhydrous ammonia (AA), urea ammonium nitrate (UAN),
and urea). In the model for anhydrous ammonia, the effect of nitrapyrin was non-significant (p
= 0.16). The model for UAN also showed a non-significant effect of enhanced efficiency
fertilizers (NBPT, NBPT+DCD, Nitrapyrin, Calcium thiosulfate, or Nutrisphere; p = 0.84). The
factor of enhanced efficiency fertilizer in the model for urea was moderately significant (p =
0.05) where the order of yield was urea+NBPT+DCD, polymer-coated urea, urea+NBPT,
urea+nitrapyrin, and finally urea, though means could not be separated statistically. The
difference in yield between urea versus urea+NBPT+DCD was 8 bushels per acre. Other factors
such as nitrogen rate, application time, seasonal precipitation and temperature, and soil
characteristics influenced corn yield to a greater extent than enhanced efficiency fertilizers. Our
understanding of the effects of enhanced efficiency fertilizers on corn yield under a variety of
growing conditions in the Midwest would be much improved by 1) a greater number of
published studies with more consistent reporting of means with variation, 2) finer scale in-
season meteorological data, 3) more complete geographic coverage, 4) simultaneous
measurement of environmental effects on water-quality and N2O emissions with
standardization of reporting in area or yield-based results. Direct modeling of NO3 and N2O was
restricted due to a lack of sufficient coverage. Overall, this meta-analysis indicates a variable
and condition-specific impact of enhanced efficiency fertilizers on corn yields in the Midwest.
The database created will be publicly available to help inform future research efforts in nitrogen
management in corn production systems.
4RM06 12/1/2015
Page 2 of 66
Table of Contents
Executive Summary ............................................................................................................... 1
Table of Contents .................................................................................................................. 2
1 Background and Motivation ........................................................................................... 3
1.1 Environmental Impacts ........................................................................................... 3
1.1.1 Subsurface Nitrate losses ............................................................................................ 3
1.1.2 Nitrous oxide emissions .............................................................................................. 4
1.1.3 Crop Yield .................................................................................................................... 5
1.2 Study Objectives ..................................................................................................... 6
2 Methods ........................................................................................................................ 6
2.1 Database Development .......................................................................................... 6
2.2 Meta-analysis vs. Systematic Review ...................................................................... 8
2.3 Meta-Analysis of crop yield through direct modeling .............................................. 8
2.3.1 Explanatory variables and random effects ................................................................. 9
2.3.2 Model construction ................................................................................................... 10
2.3.3 Weighting .................................................................................................................. 10
2.3.4 Publication bias ......................................................................................................... 11
3 Results and Discussion ................................................................................................. 11
3.1 Crop yield ............................................................................................................. 11
3.1.1 Anhydrous Ammonia ................................................................................................ 11
3.1.2 UAN ........................................................................................................................... 12
3.1.3 Urea ........................................................................................................................... 12
3.1.4 NUE ........................................................................................................................... 13
3.2 Discussion ............................................................................................................ 14
3.3 Systematic Review of Environmental Impacts ....................................................... 16
3.3.1 N2O Emissions ........................................................................................................... 16
3.3.2 Nitrate leaching ......................................................................................................... 18
4 Conclusion ................................................................................................................... 20
4.1 Recommendations for future work ....................................................................... 20
References .......................................................................................................................... 20
Appendix: Statistical Analysis .............................................................................................. 25
Page 3 of 66
1 Background and Motivation
Nitrogen management has remained at the forefront of agronomic and environmental issues
surrounding corn production in the Midwest due to increasing pressure to meet crop yield
demands while minimizing environmental impacts. Over 47 of the 87 million acres of corn
harvested in the US in 2013 (USDA NASS, 2013) were harvested in the Midwest (Illinois, Iowa,
Indiana, Ohio, Missouri, Michigan, Minnesota, and Wisconsin). This magnitude of corn
production and its high nitrogen requirement can lead to nitrate (NO3-) contamination in
groundwater and in surface waters as well as nitrous oxide (N2O) emissions to the atmosphere.
Therefore, corn production in the Midwest is arguably the most important agricultural system
for studying ways to more efficiently manage nitrogen fertilizers.
Nutrient stewardship in production agriculture through the 4Rs (right source, right rate, right
time, and right place) provides a framework for contextualizing best management practices.
Enhanced efficiency fertilizers (EEFs) including polymer coated fertilizers (PCFs) and
amendments such as nitrification inhibitors and urease inhibitors represent nitrogen
management tools that may reduce environmental impact and increase crop yields. The
exceedingly large number of possible combinations of nitrogen source, rate, time, and
placement, even within one cropping system, can make it difficult to compare results from
studies located in different regions, with variable climate, soils, and accepted management
practices. To prevent needless duplication and suggest future study directions, the soil fertility
and fertilizer community needs to periodically systematically compile what is known in order to
move forward in the most efficient manner possible. Systematic review and meta-analysis
provide a framework to compile and understand the effects of enhanced efficiency fertilizers
across the Midwestern region on environmental impacts and crop yield.
1.1 Environmental Impacts
Protecting nitrogen from transformation and subsequent losses is highly environmentally
dependent. Agricultural management practices with the appropriate EEFs may be one means by
which to reduce environmental impacts such as nitrate leaching and soil nitrous oxide
emissions.
1.1.1 Subsurface Nitrate losses
Nitrate contributes to N-leaching losses due to its negative charge and greater likelihood for
off-site movement (Jacinthe et al., 1999). Leaching of nitrogen into groundwater, subsurface
drainage, and surface runoff can result in eutrophication and is a well-documented
Page 4 of 66
environmental consequence of production agriculture (Carpenter et al., 1998; McIsaac et al.,
2001; Gardner and Drinkwater, 2009).
Nitrification inhibitors slow the conversion of ammonium to nitrate, thereby potentially
reducing losses to leaching (Wolt, 2004). Most corn hybrids take up about 40% of N applied at
preplant during the grain-filling stage (Ciampiti and Vyn, 2011) and so maintaining more N in
the soil profile as ammonium during this period of crop growth may improve efficient N use, in
addition to restriction of N losses from the soil profile.
The most commonly used nitrification inhibitors include nitrapyrin and dicyandiamide (DCD).
Nitrapyrin (sold as N-Serve® and Instinct®, Dow Agrosciences, LLC) has traditionally been added
during fall anhydrous ammonia application to prevent subsurface losses (Randall et al., 2003),
while the more recently available micro-encapsulated liquid version of nitrapyrin (Instinct)
allows for increased use of nitrification inhibitors with liquid UAN applications. Nitrification
inhibitors have been found to decrease nitrate leaching by 16% and increase yields by 7% in
studies located primarily in the Midwest (Wolt, 2004). DCD is a water-soluble nitrification
inhibitor with its effectiveness depending on temperature (declines from 10°C to 30°C) and soil
texture (McCarty and Bremmer 1989). It can be applied alone or together with the urease
inhibitor n-butyl thiophosphoric triamide (NBPT) and sold as AgrotainPlus® and SuperU® (Koch
Agronomic Services, LLC). When used alone with ammonium sulfate, DCD was found to
decrease nitrate leaching by 30% in corn grown in a Mediterranean climate (Diez et al., 2010).
Polymer-coated urea (PCU), commonly sold as “environmentally smart nitrogen” or ESN®
(Agrium, Inc.) is designed to help increase crop uptake and reduce nitrate leaching losses by
allowing a more gradual release of urea to match crop demand compared to conventional urea
(Nelson et al., 2009). In potato crops, ESN significantly reduced nitrate leaching when
compared to a split application of soluble N (Wilson et al., 2010), but was found to be less
effective in reducing nitrate leaching under corn compared to urea impregnated with
NBPT+DCD (SuperU) and to split-urea applications (Maharjan et al., 2014).
The potential for urease inhibitors to reduce nitrate leaching is not well studied. Because the
urease inhibitor NBPT is often formulated with the nitrification inhibitor DCD, few studies have
investigated the effect of NBPT alone on nitrate leaching. Urease inhibitors would not be
expected to have a great effect on nitrate leaching, but there is little empirical data to support
this assumption.
1.1.2 Nitrous oxide emissions
Nitrous oxide is about 310 times more potent than CO2 in global warming potential. While the
Agriculture sector is only responsible for 8.1% of total U.S. greenhouse gas emissions,
Page 5 of 66
agricultural activities are estimated to contribute 69% (247.2 Tg CO2 eq.) of the total 356.9 Tg
CO2 eq. anthropogenic N2O emissions in the U.S. Of this, 26% can be attributed to synthetic
fertilizer inputs (USEPA, 2013). Soil emissions of N2O have been shown to increase
exponentially with increasing nitrogen application rates, especially when N rates exceed plant
uptake capacity (Bouwman et al., 2002; Kim et al., 2012; Millar et al., 2010; Van Groenigen et al,
2010). In addition to direct N2O emissions, attempts are being made to measure and better
estimate the contributions that volatilized ammonia and leached nitrate found in estuaries,
rivers, and lakes may have in off-site N2O formation. In 2011, it was estimated that indirect N2O
emissions were 40.3 Tg CO2 eq. for cropland, with 14 Tg CO2 eq. attributed to ammonia
volatilization and 26.4 Tg CO2 eq. attributed to nitrate leaching (USEPA, 2013).
The potential of EEFs to directly or indirectly mitigate N2O emissions is based on how well
microbial processes, in interaction with environmental factors and management practices, can
be influenced to release nitrogen in conjunction with plant uptake. An understanding of which
environmental and management factors have the most effect on N2O emissions and how these
factors affect the mitigation potential of an EEF is necessary to make site specific and effective
choices. Previous meta-analyses and evaluations have looked at the effect of EEFs on N2O
emissions (Akiyama et al., 2009; Decock, 2014; Wolt, 2004), but a meta-analysis of the effects
of EEFs on environmental effects with a focus on the Midwest has yet to be accomplished.
Nitrification inhibitors have been shown to consistently and significantly reduce N2O emissions
by 38% (Akiyama et al., 2009) and 51% (Wolt, 2004), while data for PCFs were more variable
and subject to environmental and management differences. Akiyama et al. (2009) found a
significant reduction of 35% in N2O emissions with PCFs but this reduction varied greatly by
region and soil type, while Decock (2014) found no N2O reduction with use of PCF compared to
urea. Decock (2014) further analyzed the effect that environmental and management practices
would have on N2O emissions in the Midwestern region and found significant effects for
precipitation, aridity, soil carbon, soil order, and irrigation, but no significant effects for N
placement, timing, tillage, or rotation. However, Decock also found that 40% of the
heterogeneity of the data could be explained by differences among agroecological regions
which encompassed the Atlantic Maritime states, Lake states, Midwest, Great Plains and
Atlantic Gulf regions. These larger-scale effects may have masked effects within regions. For
this reason, it may be more useful to narrow the scope to within-region review and analysis to
more accurately inform best management practices and inform the agricultural community.
1.1.3 Crop Yield
Appropriate fertilizer N rate, timing, placement, and source can influence optimal nitrogen use
by crops. Of these factors, attaining optimal fertilizer rate is the most important factor for
limiting N losses (Power and Schepers, 1989), and may contribute to the variable yield
Page 6 of 66
responses measured with EEFs. Variable response in crop yield to the use of nitrification
inhibitors, urease inhibitors, and polymer-coated fertilizers is often attributed to the timing,
quantity, and frequency of rainfall after fertilizer application, as well as soil texture (Nelson et
al., 2008).
Ammonia (NH3) volatilization can lead to significant losses of N from urea-based fertilizers,
leading to yield decline. Soil pH, CEC, soil moisture, temperature, and surface residue have been
shown to be important factors influencing NH3 emissions following N application (Nelson, 1982,
Sommer et al., 2004). Urease inhibitors discourage NH3 losses to volatilization by reducing the
amount of NH3 released as a result of urea hydrolysis by inhibiting the urease enzyme. The
most commonly used urease inhibitor, NBPT, has also been shown to increase yields when used
with urea and UAN, especially in surface applications (Hendrickson, 1992). The combined effect
of nitrification and urease inhibitors may also have an additive effect in reducing NH3
volatilization compared to use of urease inhibitors alone (Zaman and Blennerhassett, 2010), but
these results have not been replicated on a wide enough range of conditions to make
generalizations of efficacy (Kim et al., 2012). Conversely, under controlled experimental
conditions a urease inhibitor (NBPT) and a nitrification inhibitor (DCD) combined with urea
caused an increase in NH3 volatilization by maintaining higher soil pH and soil ammonium
(NH4+) for greater duration, thereby offsetting the benefits of the urease inhibitor in a lab
setting (Soares et al., 2012). Further study of the indirect formation of N2O from NH3
volatilization is needed to better assess the mitigation potential of these inhibitors.
1.2 Study Objectives
For this project, our purpose was to determine whether there are environmental or agronomic
benefits to using enhanced efficiency fertilizers for N management in Midwestern corn
production systems using a systematic review of recent literature and meta-analysis of
reported results when enough information was available. Comparisons were made between
traditional nitrogen fertilizers and their respective EEFs (nitrification inhibitors, urease
inhibitors, and/or polymer-coated fertilizers) for effect on 1) agronomic yield, 2) N2O emissions,
and 3) nitrate leaching.
2 Methods
2.1 Database Development
The ability to perform a meta-analysis is dependent on the quantity and quality of data in the
literature. Studies were identified using online searches for combinations of terms such as:
Page 7 of 66
corn, maize, Midwest, enhanced efficiency fertilizers (both brand names and chemical names),
and state names. The full list can be found in the search terms tab of the database
spreadsheet. Searches were performed using Web of Science and Google Scholar, and
additional sources were collected from literature references and personal communications.
Peer-reviewed studies, scientific reports, and theses and dissertations were considered
acceptable if the study reported sufficient information. For inclusion in the dataset, studies
needed to (i) address nitrogen EEFs in Midwest corn cropping systems, (ii) be conducted under
field conditions, (iii) be published after 1994, and (iv) have reported replication and means.
Data needed to be reported with enough information to convert reported results to common
forms that allowed for inter-study comparisons (i.e. percent moisture of grain yield).
The search criteria for the systematic review and meta-analysis also included that studies take
place in the Midwestern region of the U.S., which was considered to be comprised of Illinois,
Iowa, Indiana, Minnesota, Missouri, Michigan, Ohio, and Wisconsin. In addition, Colorado,
North Dakota, Kansas, and Nebraska were included in our search and dataset to provide
additional data and to reflect a widening of the Corn Belt into the north and western regions
where corn production has increased over the years.
We recorded the following variables describing physical and chemical properties of the soil:
Soil series, soil taxonomic classification, soil order, texture class, texture group, soil organic
matter (SOM), soil organic carbon (SOC), pH, pH measurement method, bulk density, water-
filled pore space (WFPS), percent sand, and percent clay. If a soil variable was not reported in a
particular study, we obtained the value of that variable from the National Resource
Conservation Service (NRCS) Web Soil Survey (Soil Survey Staff, 2014).
The climate variables mean annual precipitation (MAP, mm) and mean annual temperature
(MAT, °C) were generated from Worldclim data that provides yearly averages of climatic
variables from the years 1950–2000 (Hijmans et al., 2005). We recorded maximum and
minimum annual temperatures by state for each study year from the beginning of the most
active planting date range to the end of the most active harvesting date range (USDA NASS,
1997). These temperatures were calculated from daily weather data collected from Daymet
(Oak Ridge National Laboratory Distributed Active Archive Center, Thornton et al., 2014).
Growing season precipitation (mm) was recorded as reported by the research paper when
available. When not reported, estimated cumulative growing season precipitation (mm) was
entered from the sum of Daymet precipitation data from the beginning of the most active
planting dates to the end of the most active harvesting dates (USDA NASS, 1997). Research site
coordinates (decimal degrees) were entered as obtained from the studies when reported or
from a search based on the name of the research site.
Page 8 of 66
Environmental impact variables included a seasonal cumulative N2O-N emissions which were area-based (kg N ha-1) and yield-scaled (g N Mg-1 grain), and flow-weighted NO3
-N leaching (mg L-1). Corn yield (Mg ha-1) was standardized to 15.5% moisture. For potential future analyses, indices of treatment effectiveness were calculated from yield and cumulative N2O-N data when available, such as: nitrogen use efficiency (NUE) and yield-scaled N2O emissions (g N2O-N Mg-1 oven dry grain):
𝑁𝑈𝐸 =𝑦𝑖𝑒𝑙𝑑 𝑜𝑓 𝑓𝑒𝑟𝑡𝑖𝑙𝑖𝑧𝑒𝑑 − 𝑦𝑖𝑒𝑙𝑑 𝑜𝑓 𝑛𝑜𝑛𝑓𝑒𝑟𝑡𝑖𝑙𝑖𝑧𝑒𝑑
𝑁 𝑟𝑎𝑡𝑒
𝑌𝑖𝑒𝑙𝑑 𝑠𝑐𝑎𝑙𝑒𝑑 𝑒𝑚𝑖𝑠𝑠𝑖𝑜𝑛𝑠 = 𝑁2𝑂 𝑒𝑚𝑖𝑠𝑠𝑖𝑜𝑛𝑠
𝑜𝑣𝑒𝑛 𝑑𝑟𝑦 𝑦𝑖𝑒𝑙𝑑
where N2O emissions are expressed as g N2O-N g ha-1 and oven dry yield is expressed as Mg ha-
1.
When information was extracted from graphical data using GraphClick (Arizona Software, 2012)
or study authors were contacted for missing information, the source of this information was
indicated in the database “source” column.
2.2 Meta-analysis vs. Systematic Review
After compiling the study data, we considered three response variables for a model-based
meta-analysis: yield, N2O emissions, and nitrate leaching. There were sufficient studies and
observations to warrant a meta-analysis of yield. However, N2O and nitrate were less well
represented in the literature. For this reason, we performed a quantitative meta-analysis on
corn yield and a systematic review of N2O emissions and nitrate leaching. Available studies are
included in the database so that as future studies are published, meta-analysis of EEF effects on
N2O emissions and nitrate leaching may be performed. For yield, our goal was to answer the
question: Do any EEFs produce significantly greater corn yield when other factors, such as
placement, application time, N rate, climate/weather, and soil characteristics are taken into
account?
2.3 Meta-Analysis of crop yield through direct modeling
We performed a meta-analysis via direct modeling of treatment means for yield. We used
multiple linear regression mixed models, specifically, analysis of covariance (ANCOVA) models,
of yield as a function of explanatory variables and random effects, under the assumption of
normality. We developed one model for each of three major nitrogen sources: Anhydrous
Page 9 of 66
ammonia (AA), urea ammonium nitrate (UAN), and urea since each fertilizer source is
characterized by different common management methods.
In each of the models, the categorical variable, or factor, fertilizer has multiple values (levels),
one of which is the non-enhanced source. The values for each model of the same N source
were: AA+nitrapyrin or AA; UAN+Agrotain, UAN+AgrotainPlus, UAN+calcium thiosulfate,
UAN+Nutrisphere, UAN+Nfusion, or UAN; and urea+Agrotain, SuperU, ESN, urea+Nutrisphere,
urea+nitrapyrin, or urea. Calcium thiosulfate (cats), Nutrisphere, and Nfusion were labeled as
“slow-release” (S.R.) in the database. If the variable fertilizer was statistically significant, then
Tukey-Kramer pairwise comparisons was used to separate the least-squares mean yield
estimates (LSmeans). The data available determined what models could be developed.
2.3.1 Explanatory variables and random effects
In addition to using fertilizer as an explanatory variable, we selected a subset of the variables
detailed in Section 2.1 for use in the statistical models that we hypothesized would have the
greatest likelihood of an effect on yield. These effects included application timing (apptime)
which was divided into fall, pre-planting (N-applied in the winter or spring > 5 days prior to
planting), at planting (N-applied 5 days before to 20 days after planting), sidedress (N-applied >
20 days after planting), pre-planting/sidedress, and at planting/sidedress. Placement (place)
was sorted into broadcast, broadcast incorporated, surface banded, and subsurface banded.
Crop rotation (rotate) had various combinations, but for modeling purposes was divided into
rotated or continuous corn cultivation. Tillage (till) was primarily either tilled or no-till, though
some instances of strip-till were included. Total N rate (ratetot) was included and transformed
to a centered-normalized form and in the form of categorical binned values (See Appendix A
2.3).
For the meteorological variables, we included the maximum and minimum daily temperatures
in the growing season and totalwater (the sum of precipitation and irrigation applied). We used
the cumulative growing season precipitation reported by the authors when available or, if none
was reported, cumulative growing season precipitation estimated using Daymet. We included
maximum and minimum temperatures to provide information about the effect an extreme
weather effects, but unfortunately, these do not provide information concerning distribution of
temperature throughout the growing season or the cumulative growing degree days, which
would have been ideal.
We included the soil variables percent sand and clay, pH, and SOM. Other variables that
described soil—soil series, soil class, soil order, and soil texture—were often were overlapping
with values of county. Thus the county random effect was used to account for between-site
differences in yield due to soil variables as well as meteorological differences.
Page 10 of 66
Functional forms were considered for continuous variables. The relationship of yield to the
continuous variables was examined as linear, quadratic, and exponential. Interactions among
the continuous variables were also considered (see Appendix A 2.4).
Random effects consisted of year nested within county to take into account variability across
locations and over years. Rotation and tillage were considered random effects in models where
there were not enough observations to test them as fixed effects due to a lack of sufficient
combinations. To implement the linear models with random effects, we used Proc Glimmix® in
SAS 9.3 (SAS Institute Inc., Cary, NC, USA).
2.3.2 Model construction
Variable selection was conducted according to a process of iterative model construction,
examining for issues of collinearity and estimability (see Section A 2.5). Residuals analysis was
performed to check model assumptions (see Section A 2.6). For each final model, we used Q-Q
plots, histograms, and boxplots of residuals to assess the assumption of normality, scatterplots
of residuals vs. predicted yield to assess heterogeneity and overall goodness-of-fit, and
scatterplots of residuals vs. each continuous variable to assess functional form decisions and
influence of individual observations on model fit.
We used the alpha = 0.01 level for the backward elimination steps. Constructing a model for
observational data entailed performing more hypothesis tests than one would perform when
analyzing the results of a designed experiment and therefore requires a more conservative
alpha level. Reducing the statistical significance level to 0.01 was implemented to reduce the
number of tests that would incorrectly support rejecting the null hypothesis of zero effect.
2.3.3 Weighting
Weighting can be an important component of a meta-analysis if individual studies differ greatly
in their levels of accuracy (Philibert et al., 2012). Each study selected for analysis will have a
different level of variability in yield for a given treatment and may also have had different
sample sizes. Weights can be assigned to observations in the model to adjust for these
attributes. For the UAN and urea models, we performed an analysis to examine to what extent
conclusions about the effect of fertilizer on yield were sensitive to the use of weights. We
considered the weights 1/SE2 (the reciprocal of the squared standard error of the mean) and r,
the number of replicates that supported each treatment mean. See Section A4.3 for a
discussion of weighting schemes and sensitivity analysis. When we used a significance level of
alpha = 0.01, we did not find that using weights changed our conclusions about the significance
of the effect of fertilizer.
Page 11 of 66
2.3.4 Publication bias
There is a well-known publication bias towards significant results which could distort the results
of meta-analysis towards finding greater significance than actually occurs (Philibert et al., 2012).
We used funnel plots to examine the likelihood of bias across all EEFs and did not find evidence
of significant publication bias (data not shown). However, through personal communications,
we are aware of many field trials involving EEFs in the Midwest that are currently unpublished.
This is often due to a reluctance of journals to publish non-significant results, often biasing the
literature towards significance.
3 Results and Discussion
3.1 Crop yield
3.1.1 Anhydrous Ammonia
The data available for anhydrous ammonia included 148 observations from 24 studies by six
different author groups in four states and six counties. The data supported a complete factorial
ANCOVA model in which the factors (and their levels) were fertilizer (AA, AA+N-serve), till (no-
till, tilled), and apptime (at planting, pre-plant, fall;
Table A 3.6). Each combination of fertilizer, till, and apptime had at least two observations. All
observations were under a corn/soybean rotation with subsurface banded placement. There
were sufficient data to test the effect of fertilizer, apptime, and till, but there were not enough
continuous corn experiments to test rotate. There were enough values of continuous variables
such as total N rate (ratetot) at each categorical variable level to test for interactions between
continuous and categorical variables. Full details regarding dataset selection and model building
can be found in Section A3.
The fixed effects in the final model included: apptime, ratetot, pH, and the interaction of ratetot
with pH (Table A 3.10). The effect of fertilizer was not significant (p = 0.16), meaning that the
1.5% difference (0.16 Mg ha-1 or 2.6 bu ac-1) between AA+nitrapyrin and AA was not significant
after controlling for other factors in the model. There were no values of apptime, till, or any of
the continuous variables that had a significant interaction with fertilizer. While a broader
literature search would be required to directly address the effect of application timing, for the
studies included in this analysis, in comparison, there was a significant 1.1 Mg ha-1 (17 bu ac-1)
increase in yield when nitrogen was applied at planting, which was significantly greater than
preplant or fall application, which were statistically similar (Table A 3.11).
Page 12 of 66
3.1.2 UAN
The final model for UAN-based fertilizers contained 262 observations from 17 studies by six
author groups in five states and 12 counties. This dataset for UAN supported an ANCOVA model
in which the factors (and their levels) were fertilizer (UAN, UAN+Agrotain, UAN+AgrotainPlus,
UAN+Calcium thiosulfate, UAN+Nutrisphere, UAN+Nfusion), apptime (at planting, at
planting/sidedress, preplant, and sidedress), and place (broadcast incorporated, broadcast,
subsurface band, and surface band). Of the 384 combinations of values of the three factors,
there were 50 that were represented by at least three observations in the dataset. Since not
every level of each factor was observed in combination with each level of each other factor, we
examined main effects of categorical variables but were not able to examine interactions
among them. Due to this lack of coverage, we chose to use rotate and till as random effects.
Additionally, there were insufficient values of the continuous variables at each level of each
categorical variable to support inferences regarding interactions between categorical and
continuous variables. Total N rate was modeled with an exponential functional form (erate) in
this model. Full details regarding dataset selection and model building can be found in Section
A4.
The fixed effects in the final model included apptime, erate, tempmax, tempmin, pH, SOM, and
the interactions between tempmax and erate, tempmin and erate, pH and erate, and SOM and
erate (Table A 4.9). Fertilizer was added back to the model to demonstrate its effect, but was
not significant (p = 0.837). None of the enhanced versions of UAN produced a significantly
higher yield than UAN alone even after controlling for other factors that affected yield, and
none of the EEFs were significantly different from any other EEF. It was not possible to examine
the interaction of fertilizer with the categorical or continuous variables. This lack of sufficient
information to form interactions is a significant gap in the literature. In comparison, application
time for the studies included in this analysis, showed that splitting nitrogen application
between at planting and sidedress increased yields between 0.9 to 16 Mg ha-1 (14 to 32 bu ac-1)
over other application times (Table A 4.10).
3.1.3 Urea
Crop yield with urea-based fertilizer sources had the greatest number of observations. The final
model was based on 479 observations from 38 studies by 16 different author groups from eight
states and 19 counties. The dataset for urea supported an ANCOVA model in which the factors
were fertilizer (urea, urea+Agrotain, SuperU, ESN, urea+Nutrisphere, urea+nitrapyrin), apptime
(at planting, at planting/sidedress, preplant, preplant/sidedress, sidedress, and fall), and place
(broadcast incorporated, broadcast, subsurface band, surface band). Of the 720 combinations
of values of the three factors, there were 74 that were represented by at least three
observations each. Since not every level of each factor was observed in combination with each
Page 13 of 66
level of each other factor, we examined main effects of categorical variables but not
interactions among them. The lack of interactions in the model means that it was not possible
to examine the effect of to apptime, place, rotate or till on fertilizer, but they could still be
included in the model as fixed effects. There were also insufficient values to support inferences
regarding interactions between continuous and categorical variables. Full details regarding
dataset selection and model building can be found in Section A4.4.
The fixed effects that remained in the final model included apptime, tempmax, tempmin, pH,
SOM, totalwater, and erate. The fixed effect of fertilizer was added back to the model to
examine its level of significance (p = 0.0525), which was not statistically significant at the 0.01
level, and therefore means for fertilizer types could not be separated statistically. The
numerical differences in LSmeans of EEFs compared to urea were at most 5%. There was a non-
significant difference of 0.5 Mg ha-1 (8 bu ac-1) for SuperU, and a 0.4 Mg ha-1 (6.4 bu ac-1) for
PCU and urea+Agrotain (Table A 4.24). None of the enhanced versions of urea produce a
significantly higher yield than urea alone, after controlling for other factors that affected yield,
and none of the EEFs was significantly different from each other. It was not possible to test the
effect of fertilizer on yield as an interaction with any of the categorical or continuous variables.
Application timing, in comparison, like in AA and UAN, had a much greater effect on yields than
fertilizer. Sidedress yielded 2.9 Mg ha-1 (46 bu ac-1) higher than preplant and 2.5 Mg ha-1 (40 bu
ac-1) higher than fall in these studies. At planting yielded 1.0 Mg ha-1 (16 bu ac-1) higher than
preplant (Table A 4.25).
3.1.4 NUE
Nitrogen use efficiency was calculated for studies that included a control (no nitrogen fertilizer
added) and is included in the database. There was a negative relationship between N rate and
NUE, showing that as increasing amounts of fertilizer are applied NUE decreased (Figure
1). Improved NUE in corn with the use of EEFs could result in N-rate reductions without
compromising yield (Shoji, 2001). When N rates were reduced in one study in Colorado, ESN
was found to improve nitrogen-use efficiency compared to urea by 4–14% at N rates from 168–
280 kg N ha-1, while no benefit to NUE was found with SuperU compared to urea (Halvorson et
al., 2014).
Additional modeling will be required to assess possible differences in effect of enhanced
efficiency fertilizers on NUE, but preliminary analysis suggests there may be differences among
fertilizer sources (data not shown). Synchronizing crop uptake with fertilizer nitrogen
availability by using EEFs may improve nitrogen use efficiency and reduce N losses, but more
information will be required to assess this hypothesis.
Page 14 of 66
3.2 Discussion
This direct modeling meta-analysis showed that, based on the reported literature, there are no
statistically significant differences between conventional fertilizer sources and their enhanced
efficiency counterparts when management, soil, and meteorological factors were taken into
account. Other meta-analyses have also found no effect of nitrification inhibitors on yield, or a
negative effect of controlled release fertilizers on yield (Quemada et al., 2013). In rice systems
however, both nitrification inhibitors and urease inhibitors were found to have a positive effect
on yield, though the effect was a modest 5.6%, while DCD was found to not be effective when
the added nitrogen was taken into account (Linquist et al., 2013). Wolt (2004) also found a
slight increase in yield (7%) with the use of nitrapyrin in a pairwise comparison, though no
meta-analytical procedure were used.
Figure 1 Nitrogen use efficiency has a significant negative relationship with N-rate applied NUE = 44.3 - 0.09*Nrate.
Page 15 of 66
The conditions under which a yield benefit with the use of EEFs may be realized are highly
variable and depend on the mechanisms for N loss. Interactions of enhanced efficiency
fertilizers with timing, application, and N rate are variable and may be heavily influenced by the
environment, thereby changing year-to-year. For example, we typically would expect to see a
reduction in N loss from a urease inhibitor if urea were broadcast applied in no-till, high residue
conditions to moist soil and followed by a dry period (Schlegal et al., 1986; Sommer et al.,
2004). For nitrification inhibitors, we might expect to see a reduction in N loss by delaying
nitrification when warm soil temperatures were combined with saturated fine-textured soils
(denitrification) or high rainfall in coarse-textured soils (leaching) (Ferguson et al., 2003).
It is possible that available data cannot appropriately describe the conditions in which N losses
lead to yield decline without the use of enhanced efficiency fertilizers. With more specific
meteorological information, it may be possible to better understand under which conditions
across the Midwest, we would be most likely to see a response to enhanced efficiency
fertilizers. Additionally, the lack of full factorial datasets for categorical variables and
insufficient coverage of continuous variables for each categorical variable for urea and UAN
models may have limited our ability to discover interactions. This gap in the literature could be
filled by new research informed by the current database compiled for this study.
It is commonly assumed that nitrogen rate is an overriding factor in the likelihood of seeing a
response to EEFs. Above optimum N rates would compensate for any savings in nitrogen due to
EEFs preventing N losses. We found no significant interactions of fertilizer and N rates, nor did
Linquist et al. (2013) in rice systems.
Our modeling results did suggest that application timing was of much greater importance in
these studies than EEFs. Though this database was not constructed to analyze application
timing across the Midwest, the relative magnitude of significance of application timing would
suggest that a meta-analysis of application timing may be a worthwhile endeavor.
We can conclude from the meta-analysis that EEFs have a rather small and mostly non-
significant effect on corn yield in the Midwest when other management and environmental
variables are taken into account. While there are certain situations in which enhanced
efficiency fertilizers can have a positive effect, when compiled across the published literature in
the Midwest, we can consider EEFs more of an insurance policy than a means to consistently
and significantly increase yield. There may be more potential benefits on the environmental
impacts, which will require even greater amounts of study.
Page 16 of 66
3.3 Systematic Review of Environmental Impacts
3.3.1 N2O Emissions
There were 11 studies and 146 total observations for N2O that fit the search criteria for this
dataset. There were seven studies with area-scaled emissions, six of which included SE. There
were nine studies where yield-scaled emissions were either given or able to be calculated, but
only one included a report of variance. A further complication with modeling N2O emissions
included the fact that measurements were collected over different intervals ranging from 89–
184 days, so that cumulative emissions were not directly comparable. There were greater
fertilizer induced emission from AA over other sources (Figure 2), but more data coverage as
studies are published and in depth modeling and analysis will be required to better examine
this effect.
Decock et al. (2014) found that variation in Fertilizer-Induced Emissions (FIE) was most affected
by N-source (44%) and the use of nitrification inhibitors (16%) and less by N timing, placement,
tillage and rotation. In comparing between fertilizer source and tillage systems, Venterea et al.
(2005) found that N2O emissions with AA were significantly higher than UAN or broadcast urea
regardless of tillage, and two out of three summary reports found the range of N2O emissions
to be higher in AA than urea and UAN (Snyder, 2009).
Nitrification inhibitors have been shown to reduce N2O emissions by 30–80% across a broad
spectrum of agricultural soils (Wolt, 2004; Akiyama et al., 2009; Decock, 2014). However,
Figure 2 Fertilizer induced N2O emissions (calculated at the percent of N applied that is lost as N2O-N) shows that AA has significantly greater emissions than any other N fertilizer treatment. Error bars represent standard error of the mean; number of observations in parentheses.
0
1
2
3
4
5
6
7
AA (3) ESN (27) Urea (32) SuperU (17) UAN (8) UAN+AT+ (6)
Fert
ilize
r in
du
ced
em
issi
on
s (%
N a
pp
lied
lost
as
N2O
)
Page 17 of 66
nitrapyrin added to AA in Iowa was not found to significantly reduce N2O emissions during the
growing season, although a significant reduction was seen between fall application and
planting (Parkin and Hatfield, 2010). In Indiana, when nitrapyrin (Instinct) was added to UAN,
N2O was reduced 19–27% over three years and significantly reduced (67%) in one year
(Omonode, 2013).
The nitrification inhibitor DCD is applied with the urease inhibitor NBPT when added to urea
(SuperU) and UAN (AgrotainPlus). In trials in Colorado, compared to urea, UAN+AgrotainPlus
and SuperU both significantly reduced N2O emissions between 46-62% (Halvorson et al. 2010,
2011, 2012). Compared to UAN, SuperU and UAN+AgrotainPlus significantly reduced N2O by 29
and 35%, respectively, in no-till and strip tilled systems (Halvorson et al., 2010).
Polymer coated fertilizers (PCFs) slow down nitrogen release but have only shown to be effective in reducing N2O emissions in some cases (Akiyama et al., 2010; Halvorson and Del Grosso, 2012). Compared to urea, ESN significantly reduced N2O emissions by 34–57%, which was higher than the reductions seen by UAN without inhibitors, although ESN and UAN were not significantly different from each other (Halvorson et al., 2010, 2012). In comparison to UAN and UAN+AgrotainPlus, ESN had significantly higher N2O emissions in Iowa (Parkin and Hatfield 2013). Venterea (2011) found no difference between ESN, SuperU, and urea regardless of tillage.
In regards to fertilizer placement, several studies have found that fertilizer banding (Argoti,
2013; Bouwman, et. al. 2002; Gagnon, 2011; Halvorson et al., 2011; Halvorson and Delgrosso,
2013) and split applications (Burzaco et al., 2013; Ma et al., 2010) increased N2O emissions,
although more data are needed to explore the differences among N placements in various
climates with different EEFs.
Increasing temperature, precipitation, and water-filled pore space (WFPS) have been shown to
increase N2O production. Higher soil temperature may increase N2O through increased
respiration with reduction in the amount of available soil oxygen and creation of anaerobic
microsites. With increasing % WFPS the ratio of NO-N/N2O-N changes, from about 3–5 in the
50–60% range to <1 at or above 80% due to consumption of NO by denitrifiers (Smith, 2003).
Pulses of N2O emissions have been observed following fertilization, soil disturbance such as
tillage, rainfall events and freeze/thaw cycles. In Illinois, Fernandez et al. (2014) found that the
largest N2O peaks were seen after heavy rainfall (>20mm) regardless of N source. Flessa (1995)
and Kaiser (1998) stated that half of all annual N2O from arable land can be generated by
freeze/thaw cycles. Halvorson (2012) found that N2O emissions were detected in plots during
the March thaw with an estimated 18, 12, 17, and 42% of the yearly N2O emissions for ESN,
SuperU, UAN+AgrotainPlus, and a no fertilizer treatment, respectively. Soil structure, carbon
content, and drainage have also been found to have significant effects on N2O formation
Page 18 of 66
(Decock, 2013; Motavalli, 2008; Maharjan, 2014; Stehfast and Bouwman, 2006). In a data
summary, N2O fluxes had been found to be higher for fine textured (2.9 kg N2O-N ha-1) than
medium textured soils (1.9 kg N2O-N ha-1) and as soil organic matter increased above 6%, N2O
emissions increased from 1.9 kg N2O-N ha-1 to 4.2 kg N2O-N ha-1 (Bouwman et al., 2002).
To better understand the complex process of N2O formation and mitigation more studies are
needed to better address the complex interactions witnessed between environmental and
management factors. Due to the high variability of the results so far obtained as to the efficacy
of EEFs, more studies comparing these products in the Midwestern region are needed,
particularly in locations that have been identified as having a high potential for N2O emissions
(i.e., high intensity and frequent rainfall events shortly after fertilization, fine-textured clayey
soils prone to poor drainage, soils with high organic matter, high N inputs, and land in transition
to no-till). Furthermore, accounting of indirect N2O emissions is needed, especially due to
potential nitrate leaching.
3.3.2 Nitrate leaching
Out of six studies and 135 total location-year observations for nitrate leaching there were only
two studies that included any measure of variability. One study reported three mean treatment
observations (Randall et al., 2011) and another reported four (Maharjan et al., 2014), both in
area-scaled results. Maharjan et al. (2014) gave 16 yield-scaled observations, but these
unfortunately could not be compared across studies. Due to the low number of studies, we
were not able to perform a meta-analysis on the effect of EEFs on nitrate leaching, but could
complete a systematic review.
In this Midwest dataset, five of the six studies that examined nitrate leaching used nitrapyrin as
a nitrification inhibitor. Some results reported flow-normalized data (kg ha-1 cm-1), which are
NO3--N losses normalized for annual flow volume and expressed on a per-centimeter
basis. Over a six-cycle corn-soybean rotation average, the addition of nitrapyrin to fall-applied
AA has been found to decrease average flow-normalized NO3--N losses by 18% and 10%
compared to fall-applied AA alone (Randall et al., 2003; Randall and Vetsch, 2005). Spring-
applied AA with nitrapyrin did not reduce flow-normalized NO3--N losses when compared to
spring-applied AA alone. However, when comparing against fall-applied AA without nitrapyrin,
spring-applied AA without nitrapyrin reduced flow-normalized NO3--N losses by 14% while
spring-applied AA with nitrapyrin reduced flow-normalized NO3--N losses by 6% (Randall and
Vetsch, 2005). In contrast, annual reports by Randall and Vetsch (2002 & 2003) indicate flow-
weighted NO3--N concentrations (mg N L-1) in drainage water under corn from fall-applied AA
were not affected by nitrapyrin. An annual report from Vetsch and Randall (2011) indicated that
adding nitrapyrin to spring-applied UAN did not affect average flow-weighted NO3--N
concentrations when compared to UAN alone. Lawlor et al. (2004) found that the addition of
Page 19 of 66
nitrapyrin to aqua-ammonia applications did not reduce average flow-weighted NO3-
concentrations in fall or spring in a corn-soybean rotation at a 168 kg ha-1 rate, and there were
no significant differences three of four years among the treatments. The results for each
application from highest NO3--N concentrations in subsurface drainage to the lowest rank as
follows: spring with inhibitor > fall with inhibitor > spring alone > fall alone, though results were
not statistically significantly different. A study by Maharjan et al. (2014) used SuperU and ESN
as treatments, resulting with the amount of NO3- leached (kg NO3
--N ha-1) under corn among
treatments as follows: ESN > SuperU > split-urea > control (no urea). The study also reported
NO3- leaching in yield-scaled units (kg N Mg-1 grain) which changed the arrangement of nitrate
leached among treatments to the following order: Control > ESN > SuperU > split-U. Reporting
N losses as yield-scaled units provides a useful way of comparing crop production versus nitrate
losses.
Subsurface losses of nitrate from continuous corn have tended to be higher than other types of
crop rotations (Weed and Kanwar, 1996). Management practices, such as diversified crop
rotations, can reduce the loss of nitrate, but there is little information regarding the
interactions of enhanced efficiency fertilizers with crop rotation. Reduced losses have occurred
in corn-soybean rotations when compared to continuous corn, though the reduction is often
dependent on climatic conditions (Randall et al., 1997). Rotations including nitrogen fixers, such
as soybeans, do not require N inputs in the legume phase of the rotation. Soybean has been
thought to contribute less to subsurface leaching of NO3--N than corn, but residual soil nitrate
from the corn phase and N that has mineralized in the soil can leach into subsurface drainage
during the soybean phase. Close agreement was found for the relative amounts of NO3--N
losses in corn-soybean rotations by Randall et al. (2003) with 45% leached in soybean and 55%
in corn, and by Randall and Vetsch (2005) with 46% leached in soybean and 54% in corn.
Insights may be gained studying the interaction of rotation with NO3--N losses, particularly
when using slow-release fertilizers, with nitrate losses in corn/soybean systems compared to
continuous corn.
Timing of N application is known to affect nitrate leaching. Depending on climatic factors and
water availability, nitrogen applied in the fall may have more opportunities for loss before crops
can utilize the nutrient than nitrogen applied in the spring. With fall application, Randall and
Mulla (2001) reported 36% more NO3--N losses in tile drainage than when compared to spring
application. Significantly higher NO3--N concentrations from fall manure applications were
found when compared to spring manure applications by Van Es et al. (2006), suggesting
increased denitrification and leaching potential from fall application. However, outside of
Minnesota there are very few studies that compare fall and spring applications of AA and their
effects on nitrate leaching with the use of inhibitors. To inform potential future regulations or
guidelines, this information would be highly desirable.
Page 20 of 66
4 Conclusion
Overall, there were non-significant differences among AA, UAN, and urea and their enhanced
efficiency fertilizer counterparts when taking into account management and environmental
factors. Application timing and N rate had much greater effects on yield than EEFs. The data
available for this direct modeling may not encompass the important information to discerning
conditions where EEFs may be most effective (i.e. rainfall and soil conditions within a few
weeks after application). A greater coverage of studies across more locations would provide a
more robust dataset for meta-analysis of the effect of local and temporal conditions on relative
yield. In regards to environmental impacts, preliminary analysis shows that fertilizer induced
N2O emissions may be greatly affected by fertilizer source, with anhydrous ammonia resulting
in greater emissions. Additional modeling will be required to evaluate this highly variable
phenomenon, published using a variety of methods of reporting conventions. Nitrate leaching,
unfortunately, shows the greatest lack of published literature despite the heavy environmental
consequences. Nitrate leaching and water-quality information reported with measures of
variability are the biggest information gap at this time for both tile-drained and non-tile-drained
systems.
4.1 Recommendations for future work
1. Standard deviations or standard errors should be estimated and reported so that results can beincluded in the meta-analysis.
2. Additional studies of corn yield in response to EEFs with urea, UAN, or AA sources as they interact with the 4Rs should be done to provide additional data points to add to the analytical procedure developed here.
3. More data should be gathered to better represent meteorological conditions (growing degree days, rainfall, humidity, soil temperature/moisture) and used in place of the surrogates used in the present analysis.
References
Akiyama, H., Yan, X., and Yagi, K. (2009). Evaluation of effectiveness of enhanced-efficiency fertilizers as mitigation options for N2O and NO emissions from agricultural soils: meta-analysis. Global Change Biology 16, 1837-1846.
Argoti, M. A. A. (2013). Nitrous oxide emissions: Measurements in corn and simulations at field and regional scale, Kansas State University. Dissertation.
Arizona-Software. GraphClick (Version 3.0) [Computer software]. 20.02.12 Awale, R., Chatterjee, A., Kandel, H., Ransom, J. (2013) Denitrification loss of fertilizer nitrogen during
drought year under soil with high clay content. Crops and Soils. July-August, 24-32. Basche, A.D., Miquez, F.E., Kaspar, T.C., and Castellano, M.J. (2014). Do cover crops increase or decrease
nitrous oxide emissions? A meta-analysis. Journal of Soil and Water Conservation 69, 471-782.
Page 21 of 66
Borenstein, M., Hedges, L.V., Higgins, J.P.T., Rothstein, H.R. (2009) Introduction to Meta-Analysis. John Wiley & Sons, Ltd. ISBN: 978-0-470-05724-7.
Bundy, L. G., and Andraski, T. W. (2007). Polymer-coated urea (ESN) for corn. In "Proc. 2007 Wisconsin Fertilizer, Aglime, and Pest Management Conf", Vol. 46, pp. 54-59. Citeseer.
Burzaco, J. P., Smith, D. R., and Vyn, T. J. (2013). Nitrous oxide emissions in Midwest US maize production vary widely with band-injected N fertilizer rates, timing and nitrapyrin presence. Environmental Research Letters 8, 035031.
Carpenter, S. R., Caraco, N. F., Correll, D. L., Howarth, R. W., Sharpley, A. N., and Smith, V. H. (1998). Nonpoint pollution of surface waters with phosphorus and nitrogen. Ecological applications 8, 559-568.
Ciampitti, I. A., and Vyn, T. J. (2011). A comprehensive study of plant density consequences on nitrogen uptake dynamics of maize plants from vegetative to reproductive stages. Field Crops Research 121, 2-18.
Decock, C. (2014). Mitigating nitrous oxide emissions from corn cropping systems in the Midwestern U.S.: potential and data gaps. Environ Sci Technol 48, 4247-56.
Diez, J., Arauzo, M., Hernáiz, P. J., Sanz, A., and Vallejo, A. (2010). Comparison of nitrification inhibitors to restrict nitrate leaching in a maize crop irrigated under mediterranean conditions. Spanish Journal of Agricultural Research 8, 481-492.
Ebelhar, S., Hart, C., and Hernandez, J. (2005). Nitrogen source and rate effects on corn in southern Illinois.
Ebelhar, S. A., C. D. Hart, J. D. Hernandez, L. E. Paul, and J. J. Warren. 2007. Evaluation of new nitrogen fertilizer technologies for corn. In E. D. Nafziger (Ed.) 2007 Illinois Fertilizer Conf. Proc., pp 64-71.
Fernández, F. G., Terry, R. E., and Coronel, E. G. (2014). Nitrous Oxide Emissions from Anhydrous Ammonia, Urea, and Polymer-Coated Urea in Illinois Cornfields. Journal of Environment Quality, 44, 415-422.
Flessa, H., Dörsch, P., and Beese, F. (1995). Seasonal variation of N2O and CH4 fluxes in differently managed arable soils in southern Germany. Journal of Geophysical Research: Atmospheres (1984–2012) 100, 23115-23124.
Foster, T. (2014). The use of nitrogen timing and nitrification inhibitors as tools in corn and wheat production in Kansas. Kansas State University. Thesis.
Ferguson, R., Lark, R., Slater, G., (2003). Approaches to management zone definition for use of nitrification inhibitors. Soil Sci. Soc. Am. J. 67, 937-947.
Gagnon, B., Ziadi, N., Rochette, P., Chantigny, M. H., and Angers, D. A. (2011). Fertilizer Source Influenced Nitrous Oxide Emissions from a Clay Soil under Corn. Soil Science Society of America Journal 75, 595.
Gardner, J. B., and Drinkwater, L. E. (2009). The fate of nitrogen in grain cropping systems: a meta-analysis of 15N field experiments. Ecological Applications 19, 2167-2184.
Gordon, W.B. (2014). Management of Urea-Containing Fertilizers for No-Tillage Corn using Nitrogen Stabilizers and Coated-Granule Technology. J. Plant Nutrition 37, 87-94.
Halvorson, A. D., and Bartolo, M. E. (2014). Nitrogen Source and Rate Effects on Irrigated Corn Yields and Nitrogen-Use Efficiency. Agronomy Journal 106, 681.
Halvorson, A. D., and Del Grosso, S. J. (2012). Nitrogen source and placement effects on soil nitrous oxide emissions from no-till corn. J Environ Qual 41, 1349-60.
Halvorson, A. D., and Del Grosso, S. J. (2013). Nitrogen placement and source effects on nitrous oxide emissions and yields of irrigated corn. J Environ Qual 42, 312-22.
Halvorson, A. D., Del Grosso, S. J., and Alluvione, F. (2010a). Nitrogen Source Effects on Nitrous Oxide Emissions from Irrigated No-Till Corn. Journal of Environment Quality 39, 1554.
Page 22 of 66
Halvorson, A. D., Del Grosso, S. J., and Alluvione, F. (2010b). Tillage and Inorganic Nitrogen Source Effects on Nitrous Oxide Emissions from Irrigated Cropping Systems. Soil Science Society of America Journal 74, 436.
Halvorson, A. D., Del Grosso, S. J., and Jantalia, C. P. (2011). Nitrogen source effects on soil nitrous oxide emissions from strip-till corn. Journal of environmental quality 40, 1775-1786.
Halvorson, A. D., Snyder, C. S., Blaylock, A. D., and Del Grosso, S. J. (2014). Enhanced-Efficiency Nitrogen Fertilizers: Potential Role in Nitrous Oxide Emission Mitigation. Agronomy Journal 106, 715.
Hatfield, J. L., and Parkin, T. B. (2014a). Enhanced Efficiency Fertilizers: Effect on Agronomic Performance of Corn in Iowa. Agronomy Journal 106, 771.
Hatfield, J. L., and Venterea, R. T. (2014b). Enhanced Efficiency Fertilizers: A Multi-Site Comparison of the Effects on Nitrous Oxide Emissions and Agronomic Performance. Agronomy Journal 106, 679.
Hendrickson, L. (1992). Corn yield response to the urease inhibitor NBPT: five-year summary. Journal of production agriculture 5, 131-137.
Hijmans, R.J., S.E. Cameron, J.L. Parra, P.G. Jones and A. Jarvis (2005) Very high resolution interpolated climate surfaces for global land areas. International Journal of Climatology 25, 1965-1978.
Jacinthe, P.-A., Dick, W., and Brown, L. (1999). Bioremediation of nitrate-contaminated shallow soils using water table management techniques: Nitrate removal efficiency. Transactions of the ASAE 42, 1251-1259.
Kaiser, E.-A., Kohrs, K., Kücke, M., Schnug, E., Heinemeyer, O., and Munch, J. (1998). Nitrous oxide release from arable soil: importance of N-fertilization, crops and temporal variation. Soil Biology and Biochemistry 30, 1553-1563.
Kim, D.-G., Hernandez-Ramirez, G., and Giltrap, D. (2013). Linear and nonlinear dependency of direct nitrous oxide emissions on fertilizer nitrogen input: A meta-analysis. Agriculture, Ecosystems & Environment 168, 53-65.
Lawlor, P.A, Baker, J.L., Stewart, M.W., Helmers, M.J. (2004). Nitrification Inhibitor and Nitrogen Application Timing Effects on Yields and Nitrate-Nitrogen Concentrations in Subsurface Drainage from a Corn-Soybean Rotation. 2004 ASAE/CSAE Annual International Meeting. Paper number 042273.
Linquist, B. A., Liu, L., van Kessel, C., van Groenigen, K.J. (2013). Enhanced efficiency nitrogen fertilizers for rice systems: Meta-analysis of yield and nitrogen uptake. Field Crops Research 154, 246-254.
Ma, B. L., Wu, T. Y., Tremblay, N., Deen, W., Morrison, M. J., McLaughlin, N. B., Gregorich, E. G., and Stewart, G. (2010). Nitrous oxide fluxes from corn fields: on-farm assessment of the amount and timing of nitrogen fertilizer. Global Change Biology 16, 156-170.
Maharjan, B., Venterea, R. T., and Rosen, C. (2014). Fertilizer and Irrigation Management Effects on Nitrous Oxide Emissions and Nitrate Leaching. Agronomy Journal 106, 703.
McCarty, G., and Bremner, J. (1989). Laboratory evaluation of dicyandiamide as a soil nitrification inhibitor 1. Communications in Soil Science & Plant Analysis 20, 2049-2065.
McIsaac, G. F., David, M. B., Gertner, G. Z., and Goolsby, D. A. (2001). Eutrophication: nitrate flux in the Mississippi River. Nature 414, 166-167.
Motavalli, P. P., Goyne, K. W., and Udawatta, R. P. (2008). Environmental impacts of enhanced-efficiency nitrogen fertilizers. Crop Management 7, 0-0.
Motavalli, P. P., Nelson, K. A., and Bardhan, S. (2012). Development of a Variable-Source N Fertilizer Management Strategy Using Enhanced-Efficiency N Fertilizers. Soil Science 178, 693-703.
Nash, P.R., Nelson, K.A., Motavalli, P.P. (2013). Corn Yield Response to Polymer and Non-Coated Urea Placement and Timings. Inter. J. of Plant Production. 7, 373-391.
Page 23 of 66
Nelson, D. W. (1982). Gaseous losses of nitrogen other than through denitrification. In "Nitrogen in agricultural soils", Vol. 22, pp. 327-363. American Society of Agronomy Madison, Wisconsin,, USA.
Nelson, K. A., Motavalli, P. P., and Dudenhoeffer, C. J. (2014). Cropping System Affects Polymer-Coated Urea Release and Corn Yield Response in Claypan Soils. Journal of Agronomy and Crop Science 200, 54-65.
Nelson, K. A., Paniagua, S. M., and Motavalli, P. P. (2009). Effect of Polymer Coated Urea, Irrigation, and Drainage on Nitrogen Utilization and Yield of Corn in a Claypan Soil. Agronomy Journal 101, 681.
Nelson, K. A., Scharf, P. C., Bundy, L. G., and Tracy, P. (2008). Agricultural Management of Enhanced-Efficiency Fertilizers in the North-Central United States. cm 7, 0.
Nelson, K. A., Scharf, P. C., Stevens, W. E., and Burdick, B. A. (2011). Rescue Nitrogen Applications for Corn. Soil Science Society of America Journal 75, 143.
Noellsch, A. J., Motavalli, P. P., Nelson, K. A., and Kitchen, N. R. (2009). Corn Response to Conventional and Slow-Release Nitrogen Fertilizers across a Claypan Landscape. Agronomy Journal 101, 607.
Norton, J.M., 2008. Nitrification in agricultural soils. In: Schepers, J.S., Raun, W.R.(Eds.), Nitrogen in Agricultural Systems. Agronomy monograph 49. American Society of Agronomy, Madison, WI, USA, pp. 173–199.
Omonode, R. A., and Vyn, T. J. (2013). Nitrification Kinetics and Nitrous Oxide Emissions when Nitrapyrin is Coapplied with Urea–Ammonium Nitrate. Agronomy Journal 105, 1475.
Parkin, T. B., and Hatfield, J. L. (2010). Influence of nitrapyrin on N2O losses from soil receiving fall-applied anhydrous ammonia. Agriculture, Ecosystems & Environment 136, 81-86.
Parkin, T. B., and Hatfield, J. L. (2014). Enhanced Efficiency Fertilizers: Effect on Nitrous Oxide Emissions in Iowa. Agronomy Journal 106, 694.
Philibert, A., Loyce, C., and Makowski, D. (2012). Assessment of the Quality of Meta-Analysis in Agronomy. Agriculture, Ecosystems, and Environment 148, 72-82.
Pittelkow, C.M., Liang, X., Linquist, B., van Groenigen, K.J., Lee, J., Lundy, M.E., van Gestel, N. Six, J., Venterea, R.T., van Kessel, C. (2015). Productivity limits and potentials of the principles of conservation agriculture. Nature 517, 365-368.
Power, J., and Schepers, J. (1989). Nitrate contamination of groundwater in North America. Agriculture, ecosystems & environment 26, 165-187.
Quemada, M., Baranski, M., Nobel-de Lange, M., Vallejo, A., Cooper, J., (2013). Meta-analysis of strategies to control nitrate leaching in irrigated agricultural systems and their effects on crop yield. Agric. , Ecosyst. Environ.174, 1-10.
Randall, G., Huggins, D., Russelle, M., Fuchs, D., Nelson, W., and Anderson, J. (1997). Nitrate losses through subsurface tile drainage in conservation reserve program, alfalfa, and row crop systems. Journal of Environmental Quality 26, 1240-1247.
Randall, G., and Vetsch, J. (2005). Nitrate losses in subsurface drainage from a corn–soybean rotation as affected by fall and spring application of nitrogen and nitrapyrin. Journal of environmental quality 34, 590-597.
Randall, G., Vetsch, J., and Huffman, J. (2003). Nitrate losses in subsurface drainage from a corn–soybean rotation as affected by time of nitrogen application and use of nitrapyrin. Journal of Environmental Quality 32, 1764-1772.
Randall, G. W., and Mulla, D. J. (2001). Nitrate nitrogen in surface waters as influenced by climatic conditions and agricultural practices. Journal of Environmental Quality 30, 337-344.
Robertson, G. P., Field, C., and Raupach, M. (2004). Abatement of nitrous oxide, methane, and the other non-CO2 greenhouse gases: the need for a systems approach. The global carbon cycle: integrating humans, climate and the natural world, 493-506.
Page 24 of 66
Rochette, P., Angers, D. A., Chantigny, M. H., and Bertrand, N. (2008). Nitrous Oxide Emissions Respond Differently to No-Till in a Loam and a Heavy Clay Soil. Soil Science Society of America Journal 72, 1363.
Schlegel, A., Nelson, D., Sommers, L., (1986). Field evaluation of urease inhibitors for corn production. Agron. J. 78, 1007-1012.
Schröder, J., Ten Holte, L., Van Keulen, H., and Steenvoorden, J. (1993). Effects of nitrification inhibitors and time and rate of slurry and fertilizer N application on silage maize yield and losses to the environment. Fertilizer research 34, 267-277.
Six, J., Ogle, S. M., Conant, R. T., Mosier, A. R., and Paustian, K. (2004). The potential to mitigate global warming with no‐tillage management is only realized when practised in the long term. Global change biology 10, 155-160.
Smith, K., Ball, T., Conen, F., Dobbie, K., Massheder, J., and Rey, A. (2003). Exchange of greenhouse gases between soil and atmosphere: interactions of soil physical factors and biological processes. European Journal of Soil Science 54, 779-791.
Smith, P., Martino, Z., and Cai, D. (2007). 'Agriculture', in Climate change 2007: mitigation. Snyder, C. S., Bruulsema, T. W., Jensen, T. L., and Fixen, P. E. (2009). Review of greenhouse gas emissions
from crop production systems and fertilizer management effects. Agriculture, Ecosystems & Environment 133, 247-266.
Soares, J. R., Cantarella, H., and de Campos Menegale, M. L. (2012). Ammonia volatilization losses from surface-applied urea with urease and nitrification inhibitors. Soil biology and biochemistry 52, 82-89.
Soil Survey Staff, Natural Resources Conservation Service, United States Department of Agriculture. Web Soil Survey. Available online at http://websoilsurvey.nrcs.usda.gov/. Accessed [8/15/2014].
Sommer, S.G., Schjoerring, J.K., Denmead, O., (2004). Ammonia emission from mineral fertilizers and fertilized crops. Adv. Agron. 82, 557-622.
Stehfest, E., and Bouwman, L. (2006). N2O and NO emission from agricultural fields and soils under natural vegetation: summarizing available measurement data and modeling of global annual emissions. Nutrient Cycling in Agroecosystems 74, 207-228.
Stehouwer, R., and Johnson, J. (1990). Urea and anhydrous ammonia management for conventional tillage corn production. Journal of Production Agriculture 3, 507-513.
Thornton, P.E., M.M. Thornton, B.W. Mayer, N. Wilhelmi, Y. Wei, R. Devarakonda, and R.B. Cook. 2014. Daymet: Daily Surface Weather Data on a 1-km Grid for North America, Version 2. Data set. Available on-line [http://daac.ornl.gov] from Oak Ridge National Laboratory Distributed Active Archive Center, Oak Ridge, Tennessee, USA. Data accessed between 2014/07/15 and 2014/11/24. Temporal range: 1987/04/26-2013/11/08. Spatial range: N=45.38, S=33.60, E=-81.33, W=-104.98. http://dx.doi.org/10.3334/ORNLDAAC/1219
USDA-NASS (1997). Usual Planting and Harvesting Dates for U.S. Field Crops (Agricultural Handbook Number 628). http://www.nass.usda.gov/Publications/Usual_Planting_and_Harvesting_Dates/uph97.pdf
USEPA (2013). Inventory of U.S. Greenhouse Gas Emissions and Sinks: 1990-2011. EPA 430-R-13-001,
http://www.epa.gov/climatechange/emissions/usinventoryreport.html. Van Es, H. M., Sogbedji, J. M., and Schindelbeck, R. R. (2006). Effect of manure application timing, crop,
and soil type on nitrate leaching. Journal of environmental quality 35, 670-679. Van Groenigen, J. W., Velthof, G. L., Oenema, O., Van Groenigen, K. J., and Van Kessel, C. (2010).
Towards an agronomic assessment of N2O emissions: a case study for arable crops. European Journal of Soil Science 61, 903-913.
Page 25 of 66
van Kessel, C., Venterea, R., Six, J., Adviento-Borbe, M. A., Linquist, B., and van Groenigen, K. J. (2013). Climate, duration, and N placement determine N2O emissions in reduced tillage systems: a meta-analysis. Glob Chang Biol 19, 33-44.
Venterea, R. T., Bijesh, M., and Dolan, M. S. (2011). Fertilizer source and tillage effects on yield-scaled nitrous oxide emissions in a corn cropping system. J Environ Qual 40, 1521-31.
Venterea, R. T., Burger, M., and Spokas, K. A. (2005). Nitrogen oxide and methane emissions under varying tillage and fertilizer management. J Environ Qual 34, 1467-77.
Venterea, R. T., Halvorson, A. D., Kitchen, N., Liebig, M. A., Cavigelli, M. A., Grosso, S. J. D., Motavalli, P. P., Nelson, K. A., Spokas, K. A., Singh, B. P., Stewart, C. E., Ranaivoson, A., Strock, J., and Collins, H. (2012). Challenges and opportunities for mitigating nitrous oxide emissions from fertilized cropping systems. Frontiers in Ecology and the Environment 10, 562-570.
Venterea, R. T., and Stanenas, A. J. (2008). Profile analysis and modeling of reduced tillage effects on soil nitrous oxide flux. Journal of environmental quality 37, 1360-1367.
Vetsch, J. A., and Randall, G. W. (2004). Corn production as affected by nitrogen application timing and tillage. Agronomy Journal 96, 502-509.
Weber, H.S. (2008). Use of Nitrogen Management Products and Practices to Enhance Yield and Nitrogen Uptake in No-Till Corn and Grain Sorghum. Kansas State University. North Central Extension-Industry Soil Fertility Conference 25.
Weed, D., and Kanwar, R. (1996). Nitrate and water present in and flowing from root-zone soil. Journal of Environmental Quality 25, 709-719.
Wilson, M. L., Rosen, C. J., and Moncrief, J. F. (2010). Effects of polymer-coated urea on nitrate leaching and nitrogen uptake by potato. Journal of environmental quality 39, 492-499.
Wolt, J. D. (2004). A meta-evaluation of nitrapyrin agronomic and environmental effectiveness with emphasis on corn production in the Midwestern USA. Nutrient Cyclcling in Agroecosystem 69, 23-41.
Yahdjian, L., and Sala, O. E. (2010). Size of precipitation pulses controls nitrogen transformation and losses in an arid Patagonian ecosystem. Ecosystems 13, 575-585.
Zaman, M., and Blennerhassett, J. D. (2010). Effects of the different rates of urease and nitrification inhibitors on gaseous emissions of ammonia and nitrous oxide, nitrate leaching and pasture production from urine patches in an intensive grazed pasture system. Agriculture, Ecosystems & Environment 136, 236-246.
Appendix: Statistical Analysis
A1 Introduction
The specific objective of this quantitative meta-analysis is to determine whether using an
enhanced efficiency fertilizer (EEF) will produce greater corn yield than will using the non-
enhanced version of the same N source, to compare the yields produced by different EEFs, and
to determine whether and to what extent the ability of EEFs to increase yield is affected by
placement, application time, application rate, or other management, meteorological, or soil
variables. To this end, we developed three models of the response variable yield—one for each
of the primary N sources anhydrous ammonia (AA), urea ammonium nitrate (UAN), and urea—
Page 26 of 66
as a function of the explanatory variables fertilizer, application timing, rate, placement,
cropping system, tillage, climate and meteorological variables, and soil characteristics.
A1.1 Notation and language use
The factors were italicized to specifically refer to the mathematical variable. In some cases,
particularly for the variables fertilizer and yield, the use of the word as a mathematical variable
is almost the same as is its normal usage, and in this case either italics or non-italics may be
used. Italics are also used for emphasis, as they are in non-mathematical documents.
Variable names refer to column headings in the spreadsheet unless the variable was created
from those columns, in which case it was be defined in the text.
A factor is a categorical variable being considered as an independent variable in a model.
The following words are synonyms: independent variable, explanatory variable, predictor
variable, and covariate.
A2 Methods
A2.1 Selecting data for models
The availability of data determined which models could be developed and what questions could
be answered with them. We started with the 1281-observation dataset found in the “yield” tab
of the database (Midwest_Corn_EEF_Meta-analysis Database.xlsx). An observation—or row—in
this dataset represents a yield treatment mean from a study with associated explanatory
variables. We excluded 11 observations from 3 studies that were missing the yield treatment
mean due to technical problems the authors encountered, but we kept the other observations
from these studies. We omitted 22 additional observations from three studies in which
treatment means were reported as averages over factors of interest (Nelson 13, Nelson 11, and
Burzaco 14 in the dataset). These studies were averaged over application timing, over three
cropping systems, and with results representing complicated averages over sites and years
(Nelson et al., 2011; Nelson et al., 2014; Burzaco et al., 2014).
Those edits produced the 1248-observation dataset with source-inhibitor-fertilizer
combinations shown in Table A 2.1, which gives the number of observations in each
combination and the number of observations for which the standard error (SE) associated with
the treatment mean was available. SE could have been reported in the paper, provided by the
author in personal communications, or derived from the standard deviation (SD) that was
Page 27 of 66
reported or provided by the author (𝑆𝐸 = 𝑆𝐷/√𝑟 ), where 𝑟 is the number of replicates
supporting the treatment mean). The total number of observations with SE is 917.
There was considerable range of number of observations for the different fertilizer source and
enhancement types (Table A 2.1). Based on the observation totals, and the fact that some
management practices are inherently different for different N sources (e.g., AA is only applied
subsurface), we estimated a separate model to assess the effects of enhancements for AA,
UAN, and urea. The other fertilizer sources were represented by too few observations to
attempt to model these effects.
Table A 2.1 Source-inhibitor-fertilizer combinations with number of observations (treatment means) and number of observations that include SE.
Combination Source Inhibitor Fertilizer n n_with_se
1 AA nitrapyrin AA+nitrapyrin 72 23
2 AA none AA 92 30
3 Aq. A nitrapyrin Aq. A+Nitrapyrin 8 0
4 Aq. A none Aq. A 16 0
5 Nitamin S.R. Nitamin 6 6
6 Nurea S.R. Nurea 1 1
7 UAN NBPT UAN+Agrotain 56 56
8 UAN NBPT+DCD UAN+AgrotainPlus 58 46
9 UAN S.R. UAN+Nfusion 2 2
10 UAN S.R. UAN+Nutrisphere 9 0
11 UAN nitrapyrin UAN+Instinct 13 13
12 UAN none UAN 117 105
13 UAN thiosulfate UAN cats 40 40
14 ammonium sulfate DCD ammonium sulfate+DCD 2 0
15 ammonium sulfate none ammonium sulfate 4 0
16 control control control 194 145
17 urea NBPT urea+Agrotain 51 40
18 urea NBPT+DCD SuperU 78 74
19 urea PCF Duration III 2 2
20 urea PCF ESN 160 118
21 urea PCF PCU 8 8
22 urea PCF urea/ESN 4 0
Page 28 of 66
Combination Source Inhibitor Fertilizer n n_with_se
23 urea S.R. urea+Nutrisphere 17 14
24 urea nitrapyrin urea+Instinct 28 28
25 urea nitrapyrin urea+N-Serve 4 0
26 urea nitrapyrin urea+nitrapyrin 1 1
27 urea none urea 205 165
1248 917
A2.2 Meta-analysis via direct modeling
We performed a meta-analysis via direct modeling of treatment means for yield. We used
multiple linear regression models of yield as a function of explanatory variables and random
effects, under the assumption of normality.
In each of the models, the categorical variable, or factor, fertilizer has multiple values (levels),
one of which is the non-enhanced N source: AA, UAN, or Urea. The other values are enhanced
versions of the same N source: AA+nitrapyrin for AA; UAN+Agrotain, UAN+AgrotainPlus,
UAN+calcium thiosulphate, UAN+Nutrisphere, UAN+Nfusion for UAN; and urea+Agrotain,
SuperU, ESN, urea+Nutrisphere, urea+nitrapyrin for urea. If the variable fertilizer was
statistically significant, then the mean yield for at least one of the fertilizer/enhancement
combinations for that N source was significantly different from the means of all the others. To
interpret a statistically significant effect of fertilizer, we looked at the pairwise comparisons of
model-based least-squares mean estimates (LS means) for each fertilizer to determine which
means were significantly different.
To determine the effect of fertilizer using a meta-analysis, it was necessary to control for factors
that might have contributed to variability in yield but may have differed between studies
covering a wide variety of sites and years. Within an individual study, researchers control for or
include variables such as rate, application time, placement, cropping system, and tillage;
meteorology, soil characteristics, drainage systems, and topology are controlled by using a
single site, by replicating treatments over multiple sites, or by blocking over individual plots on
the same site that have different characteristics.
Meta-analysis data, however, is observational rather than produced via experimental design.
For this analysis, factors other than fertilizer that might have affected yield were controlled by
including each in the model as either a fixed effect or a random effect. We use a fixed effect
when we were interested in drawing conclusions about the factor itself and its interaction with
fertilizer, and if the data are sufficient to support such inferences. We use a random effect if we
Page 29 of 66
were not interested in quantifying the effect of the factor itself, but must account for its effect
to avoid drawing erroneous conclusions about the effect of fertilizer, or if the data were
insufficient for quantification as a fixed effect. It was necessary to include all factors that might
affect yield simultaneously in a single model, so that the degree to which each factor and/or
each interaction among factors describing variability in yield could be appropriately discerned
and allocated. Constructing multiple models of yield using only one explanatory factor at a
time, which has been done in other meta-analyses (Basche, et al. 2014, Decock 2014), can lead
to inflated estimates of effect of these factors on yield and their level of significance because
each factor is the only one being allowed to explain variability in the model, without competing
with other factors.
A2.3 Explanatory variables and random effects
To control for the simultaneous effect of the several factors examined in this study, we selected
a subset of the variables detailed in Section 2.1 for use in the statistical models. Of the 4R’s, the
N source is accounted for by using a different model for AA, UAN, and Urea. Application timing,
or apptime, was divided into fall, pre-planting (N-applied in the winter or spring > 5 days prior
to planting), at planting (N-applied 5 days before to 20 days after planting), side dress (N-
applied > 20 days after planting), pre-planting/sidedress, and at planting/sidedress. Placement
or place was sorted into broadcast, broadcast incorporated, surface banded, and subsurface
banded. Other management variables included crop rotation, which was represented by the
variable rotate, which takes the value “no” for continuous corn, and “yes” otherwise. The
variable till takes values “tilled” or “no-till.”
Total N rate (ratetot) was a continuous variable. Since the positive effect of N rate on yield is
known to diminish with greater yield, we also considered use of a transformation of ratetot that
would produce such a curve: erate = exp(-(ratetot – mean)/standard deviation). The mean and
standard deviation of ratetot were calculated separately for each model. For the UAN and urea
models, we also considered ratetot as a categorical variable describing discrete bins of N rate
during the exploratory model-building process. For the AA model, we found a linear function of
ratetot to be sufficient. This finding is consistent with the fact that any exponential function can
be approximated by a line over a small range of values, and the range of values of ratetot for
the AA model was much smaller (90-194 kg ha-1) than that of UAN (67-268 kg ha-1) and Urea
(56-336 kg ha-1).
We included the meteorological variables tempmax and tempmin, the maximum and minimum
daily temperatures in the growing season, and totalwater (mm), the sum of precipitation
(precip) and irrigation applied (irrigapp). For precip we used the cumulative growing season
precipitation reported by the authors (preciprep) or, if none was reported, cumulative growing
Page 30 of 66
season precipitation estimated using Daymet (precipest). Maximum and minimum
temperatures provide information about the effect of short-term extreme temperatures might
have on yield, but do not give information about the distribution of temperature throughout
the growing season or the cumulative growing degree days. Cumulative precipitation, while
important, does not show the distribution of precipitation over the growing season. To account
for meteorological differences between sites, and between years at one site, that might affect
yield but are not quantified by variables in our data, we include a random effect of year nested
within county.
We included the site- and/or plot-specific variables sand, the percent of the soil that is sand;
clay, the percent of the soil that is clay; pH; and soil organic matter (SOM). We excluded other
variables that described soil (soil series, soil class, soil order, and soil texture) due to the fact
that their values often were confounded with values of county. Thus the county random effect
accounts for between-site differences in yield due to soil variables not modeled explicitly and
other site-specific variables, in addition to the meteorological differences.
A2.4 Functional forms
To determine the functional form (linear, quadratic, or exponential) through which each of the
continuous variables (ratetot, tempmax, tempmin, totalwater, sand, clay, pH, and SOM) should
enter the model, we examined scatterplots of yield vs. each variable, existing knowledge of the
nature of the relationship between yield and these variables, and plots of residuals vs. each
variable after fitting a model with a linear functional form for the variable.
We considered interactions among the continuous variables. It might have been the case, for
example, that the effect of tempmax was be different for a site/year with lower totalwater than
for one with higher totalwater.
Creating higher-order polynomial terms and interactions can result in a design matrix with
multicollinearity, which can lead to inflated standard error estimates for regression parameters,
inflated p-values, and erroneously declaring terms to be non-significant. To minimize
multicollinearity, we centered and scaled each continuous variable—by subtracting its mean
and dividing by its standard deviation—prior to creating interaction terms, quadratic terms, or
exponential terms. Centering minimizes the degree to which creation of interactions and higher
order terms produces multicollinearity, while scaling mitigates other issues in the numerical
algorithms that can occur when very large numbers occur in the same matrix as very small
numbers. Because all models were run on centered-and-scaled variables, the regression
parameter estimates must be interpreted with care, but interpreting regression parameter
estimates is not the objective of this study.
Page 31 of 66
A2.5 Variable selection
Variable-selection decisions occurred at many steps along the way. The first stage was in the
compilation of the database. The second was in selecting that subset of each type of variable—
management practice, climate, meteorology, or soil—that best supported a regression model,
as listed in Section A2.2. It is also necessary to eliminate from consideration some variables due
to collinearity and estimability issues.
A2.5.1 Collinearity
At a given research site, continuous variables such as sand, clay, SOM, and sometimes pH are
constant over all treatment combinations and years. The values of tempmax, tempmin, and
totalwater vary year-to-year at a given site, but are the same for all treatment means at the
same site in the same year. When continuous variables are measured at a small number of
values, sometimes putting many of them together in a single model with interactions and
higher order terms results in a great deal of multicollinearity, even after using prophylactic
measures such as centering and scaling.
To assess and eliminate remaining collinearity issues, we used an eigenvalue/eigenvector
analysis of the design matrix for which the columns consist of all (centered-and-scaled)
continuous variables and their pairwise interactions. Such an analysis captures overall
collinearity among all columns in the design matrix, as opposed to comparing correlation
coefficients of two variables at a time. This analysis gives the eigenvalues of X’X, where X is the
design matrix, along with the condition index for each eigenvector. The largest condition index
is called the condition number. Any condition index over 30 indicates the presence of a
moderate collinearity. If it is over 100, it is a severe collinearity. The analysis also gives the
percent of variability in each column explained by the eigenvector associated with each
condition index. This allows identification of the columns involved in the collinearity. We
identified interaction terms that contributed to collinearities and eliminated them until the
condition number was around 30. We did not eliminate main effects in the collinearity analysis.
Note that if an interaction term was eliminated based on collinearity, the variability in yield that
might be attributable to that interaction will be accounted for by other variables that remain,
by the definition of collinearity. Thus eliminating variables in this collinearity exercise does not
prevent us from using variables other than fertilizer as controls.
A2.5.2 Estimability
Limitations in what variables can be considered also occur when not every combination of two
categorical variables is represented in the data. Section A3 shows that the AA data support a
full-factorial model since every combination of the categorical variables (apptime, till, and
fertilizer) is represented in the data. For UAN and Urea, however, (Sections A4.1.4 and A4.5.4)
Page 32 of 66
the categorical variables must be considered as either main effects only or each 3- or 4-way
interaction must be treated as the value of a new categorical variable we could call treatment.
In other words, for UAN and urea, we could consider main effects of the categorical variables or
interactions, but not both, in the same model because of lack of balance in the data.
For the UAN and urea models, we also were unable to consider 2-way interactions between
continuous and categorical variables due to not having enough different values of a continuous
variable for each value of the categorical variables. A concrete example of this phenomenon is
given in Section A4.2.3.
A2.6 Residual analyses
For each final model, we used Q-Q plots, histograms, and boxplots of residuals to assess the
assumption of normality, scatterplots of residuals vs. predicted yield to assess heterogeneity
and overall goodness-of-fit, and scatterplots of residuals vs. each continuous variable to assess
functional form decisions and influence of individual observations on model fit.
A3 Anhydrous ammonia (AA)
A3.1 Selecting data
A3.1.1 Original AA data
Since there was only one enhanced efficiency fertilizer in the AA data (Table A 3.1), the variable
fertilizer took only two values: AA+nitrapyrin and AA. The number of observations for each
fertilizer and the overall number of observations was sufficient to support modeling over a
variety of other conditions. If we had omitted observations for which SE was missing, more than
two-thirds of the data would have been unusable.
Table A 3.1 Source-inhibitor-fertilizer combinations in original AA data, number of observations, and number of observations with SE.
Combination Source Inhibitor Fertilizer n n_with_se
1 AA nitrapyrin AA+nitrapyrin 72 23
2 AA none AA 92 30
Total 164 53
Page 33 of 66
A3.1.2 Edits
The numbers of studies, author sets, states, counties, and years covered by the original AA
data for all observations and for only observations including SE are shown in
Table A 3.2. To draw conclusions based on observations covering as wide a variety of
conditions as possible, for the AA model we included all data, not omitting those observations
lacking SE.
Table A 3.2 Numbers of studies, author sets, states, counties, and years covered by the original AA data, for all observations and for observations that included SE.
Studies Author sets States Counties Years
All observations 26 6 5 7 27
Observations with SE 20 3 3 2 10
To be included in the dataset for the model, values of explanatory variables such as rate,
placement, application timing, tillage, and rotation needed to be adequately represented and
distributed among observations for AA vs. AA+nitrapyrin.
At planting/sidedress and pre-planting/sidedress applications are only found with AA, not with
AA+nitrapyrin, in the original AA data (Table A 3.3). To avoid potential bias in conclusions about
the effect of fertilizer vs the effect of apptime, we omitted these seven observations from the
data.
Table A 3.3 Apptime-fertilizer combinations in the original AA data, number of observations, and number of observations with SE.
Combination apptime Fertilizer n n_with_se
1 AtPl AA 15 6
2 AtPl AA+nitrapyrin 11 4
3 AtPl/SD AA 2 0
4 PrePl AA 38 16
5 PrePl AA+nitrapyrin 20 9
6 PrePl/SD AA 5 0
7 fall AA 32 8
8 fall AA+nitrapyrin 41 10
164 53
Only nine observations in the AA data were continuous corn, which was not enough to allow us
to discern effects of rotation (
Page 34 of 66
Table A 3.4). We omitted these observations from the data, which ensured that the cropping
system is controlled in the data.
Table A 3.4 Rotate-fertilizer combinations in the original AA data, number of observations, and number of observations with SE. Rotate = no represents continuous corn, and rotate = yes includes corn/soy, corn /double-crop soy, and corn/red clover.
Combination rotate Fertilizer n n_with_se
1 no AA 7 5
2 no AA+nitrapyrin 2 0
3 yes AA 85 25
4 yes AA+nitrapyrin 70 23
164 53
Strip-tilled plots were represented for AA but not for AA+nitrapyrin (Table A 3.5). To avoid
confounding of the effect of fertilizer with the effect of till, we omitted those two observations.
Table A 3.5 Till-fertilizer combinations in the original AA data, number of observations, and number of observations with SE.
Combination till Fertilizer n n_with_se
1 no-till AA 34 0
2 no-till AA+nitrapyrin 20 0
3 strip AA 2 2
4 tilled AA 56 28
5 tilled AA+nitrapyrin 52 23
164 53
Final AA data
Table A 3.6 shows the combinations of values of apptime (AtPl = at planting, PrePl = pre-plant,
fall), till (no-till, tilled), and fertilizer (AA, AA+nitrapyrin) that remained after the above edits.
The observations omitted in each step above were not mutually exclusive, thus a total of 16
observations were omitted, resulting in a 148-observation dataset that allowed a complete
factorial model to be fit. All combinations of each level of each factor were represented, with at
least two observations for each combination. The dataset included only corn/soybean rotations
and subsurface banded placements.
Page 35 of 66
Table A 3.6 Apptime-till-fertilizer combinations in the final AA data, number of observations, and number of observations with SE.
Combination apptime till Fertilizer n n_with_se
1 AtPl no-till AA 4 0
2 AtPl no-till AA+nitrapyrin 2 0
3 AtPl tilled AA 11 6
4 AtPl tilled AA+nitrapyrin 9 4
5 PrePl no-till AA 9 0
6 PrePl no-till AA+nitrapyrin 4 0
7 PrePl tilled AA 23 11
8 PrePl tilled AA+nitrapyrin 15 9
9 fall no-till AA 12 0
10 fall no-till AA+nitrapyrin 12 0
11 fall tilled AA 19 8
12 fall tilled AA+nitrapyrin 28 10
Total 148 48
Fertilizer rate was balanced reasonably over the values of fertilizer (Figure A 3.1). The
histograms below show that most values of ratetot (total N rate) for AA and AA+nitrapyrin were
in the 114-150 kg ha-1 bins. AA had six, and AA+nitrapyrin five, observations in the 174 kg ha-1
bin or higher. AA+nitrapyrin had four observations in the 90 kg ha-1 bin. If AA+nitrapyrin truly
produced higher yield than AA, the small number of observations for AA with higher values of
ratetot than those for AA+nitrapyrin was unlikely to skew results toward declaring AA to
produce higher yield. The same was true for the four observations for AA+nitrapyrin that fell in
the 90 kg ha-1 bin, which was lower than that for any AA observation. After the final model was
fit, we examined influence statistics to see if any observations exerted much more influence
than others on conclusions about effects of different explanatory variables on yield. None of
the highest and lowest values of ratetot had unusual influence on conclusions about the effect
of ratetot on yield.
Page 36 of 66
Figure A 3.1 Histograms of ratetot where fertilizer = AA and AA+N-Serve.
We also examined balance of the other continuous variables between the two levels of
fertilizer, and all of those variables had better balance than did ratetot. We further examined
balance of all continuous variables over levels of apptime and till, and again saw good balance.
A3.2 Model runs
A3.2.1 Categorical variable model
To see if the conclusions about the effect of fertilizer on yield were sensitive to the inclusion of
the continuous variables, we began by fitting a categorical-variable-only AA model, with factors
fertilizer, apptime, and till. The model included a random effect of year nested within county.
Ratetot will be included as a continuous variable in model runs described below.
Because all combinations of all levels of each of the three factors are represented, it was
possible to examine main effects of each factor, along with two- and three-way interactions.
Table A 3.7 gives the tests of significance for the full model. None of the terms containing
fertilizer--the main effect or the two- or-three-way interaction terms— were statistically
significant.
------------------- YLD AA -------------------
26
5
26
14
13 3
0
10
20
30
40
Perc
ent
AA
4
25
5
24
75
90 102 114 126 138 150 162 174 186 198
0
10
20
30
40
Perc
ent
AA
+N
-Serv
e
ratetot
Page 37 of 66
Table A 3.7 Type III tests of fixed effects for the anhydrous ammonia categorical variable only model. The values of the variable fertilizer are AA and AA+nitrapyrin.
Effect
Num
DF
Den
DF F Value Pr > F
Fertilizer 1 103.3 2.34 0.1292
apptime 2 105 9.09 0.0002
till 1 33.92 5.15 0.0297
apptime*Fertilizer 2 103.3 0.77 0.4655
till*Fertilizer 1 103.3 0.38 0.5391
apptime*till 2 105 1.31 0.2746
apptime*till*Fertili 2 103.3 0.32 0.7251
Terms were backward eliminated until all terms were significant at the alpha = 0.05 level, which
resulted in a model that included only the main effects of apptime (p-value <0.0001) and till (p-
value 0.035). Further elimination to the alpha = 0.01 and 0.001 levels resulted in a model that
included only the main effect of apptime. Table A 3.8 and Table A 3.9 give the LSMeans for till
and apptime in the alpha = 0.05 model. Mean yield for tilled plots, 11 Mg ha-1 is significantly
higher than for no-till plots, 9.8 Mg ha-1. Mean yield for at planting applications was significantly
higher than for pre-planting and fall applications, which were not significantly different from
each other.
Table A 3.8 Tukey-Kramer grouping for till least squares means (Alpha=0.05). LSmeans with the same letter are not significantly different.
till Estimated mean yield (Mg ha-1)
tilled 11.5 A
no-till 9.8 B
Table A 3.9 Tukey-Kramer grouping for apptime least squares means (Alpha=0.05). LSmeans with the same letter are not significantly different.
apptime Estimated mean yield (Mg ha-1)
AtPl 11.3 A
PrePl 10.3 B
fall 10.3 B
Page 38 of 66
A3.2.2 Categorical and continuous variable models
We next put the factors fertilizer, apptime, and till into a model along with all of the
continuous variables, two-way interactions between each of the continuous variables and each
of the three categorical factors, and those two-way interactions between pairs of continuous
variables that remained after the collinearity analysis. For this model, fertilizer occurred as a
main effect, in two-way interactions with each of apptime and till, in a three-way interaction
with apptime and till, and in interactions with each of the continuous variables, ratetot,
tempmax, tempmin, sand, clay, pH, SOM, and totalwater. None of the p-values for any of the
terms containing fertilizer was lower than 0.26 (not statistically significant).
A3.2.3 Final model
We back-eliminated from this model all terms that were not significant at the alpha = 0.01 level.
Main effects remained in the model if there was a significant interaction of that effect with
another variable. The final remaining variables included apptime, ratetot, pH, and ratetot*pH.
To this model we added fertilizer, and it had a non-significant p-value of 0.16 (Table A 3.10).
Thus the effect of fertilizer on yield was not significant, and this conclusion was not sensitive to
variable-selection decisions or sequence.
Table A 3.10 Type III tests of fixed effects for the AA final model.
Effect
Num
DF
Den
DF F Value Pr > F
Fertilizer 1 107.6 2.00 0.1604
apptime 2 109.6 17.18 <.0001
ratetot 1 106.5 0.88 0.3496
pH 1 140.1 0.21 0.6463
ratetot_ph 1 113.4 8.65 0.0040
Mean yield for at-planting applications was significantly higher than those for pre-planting and
fall applications, which were not significantly different from each other (Table A 3.11).
Table A 3.11 Tukey-Kramer grouping for apptime least squares means (alpha=0.05). LSmeans with the same letter are not significantly different.
apptime Estimated mean yield (Mg ha-1)
AtPl 11.7 A
PrePl 10.7 B
fall 10.6 B
Page 39 of 66
A4 Urea ammonium nitrate (UAN)
A4.1 Selecting data
A4.1.1 Original UAN data
UAN had 6 enhanced efficiency fertilizers to examine. We omitted UAN + Nfusion from the
analysis due to having too few observations.
Table A 4.1 Source-inhibitor-fertilizer combinations in the original UAN data, number of observations, and number of observations with SE.
Combination Source Inhibitor Fertilizer n n_with_se
1 UAN NBPT UAN+Agrotain 56 56
2 UAN NBPT+DCD UAN+AgrotainPlus 58 46
3 UAN S.R UAN + Nfusion 2 2
4 UAN S.R UAN+Nutrisphere 9 0
5 UAN nitrapyrin UAN+Instinct 13 13
6 UAN none UAN 117 105
7 UAN thiosulfate UAN+cats 40 40
Total 295 262
A4.1.2 Edited UAN data
Table A 4.2 gives the remaining number of observations for each fertilizer. The factor fertilizer
had 6 levels: UAN+Agrotain, UAN+AgrotainPlus, UAN+Nutrisphere, UAN+Instinct, UAN, and
UAN+cats. If we included observations that are missing SE, we gained ten observations for
UAN+AgrotainPlus, nine for UAN+Nutrisphere, and 12 for UAN.
Table A 4.2 Values of the variable fertilizer in the edited UAN data, number of observations, and number of observations with SE.
Fertilizer n n_with_se
UAN+Agrotain 56 56
UAN+AgrotainPlus 58 46
UAN+Nutrisphere 9 0
UAN+Instinct 13 13
Page 40 of 66
Fertilizer n n_with_se
UAN 117 105
UAN+cats 40 40
Total 293 260
The factor apptime had 4 levels: at planting (AtPl), at planting/ sidedress (AtPl/SD), pre-planting
(PrePl), and sidedress (SD) (Table A 4.3). Including observations with missing SE added 21 and
12 additional observations for at planting and sidedress, respectively.
Table A 4.3 Values of the variable apptime in the edited UAN data, number of observations, and number of observations with SE.
apptime n n_with_se
AtPl 211 190
AtPl/SD 16 16
PrePl 27 27
SD 39 27
The factor place had 4 levels in the UAN data: broadcast, broadcast incorporated (bcast inc),
subsurface banded (sub band), and surface banded (surf band) (Table A 4.4). Including
observations with missing SE improved coverage of broadcast, subsurface banded, and of
surface banded.
Table A 4.4 Values of the variable place in the edited UAN data, number of observations, and number of observations with SE.
place n n_with_se
bcast inc 112 112
broadcast 134 114
sub band 23 21
surf band 24 13
The continuous variable ratetot ranged in value from 67 to 268 kg ha-1. To allow a comparison
of coverage similar to that for the other 4R factors, we created a categorical variable rate to
divide ratetot into four evenly spaced bins from 67-117, 118-167, 168-217, and 218-268 kg ha-1
(
Table A 4.5). Including observations with missing SE was especially helpful for rates between 67
and 117 kg ha-1, creating an increase from 65 to 92 observations.
Page 41 of 66
Table A 4.5 Values of the variable rate in the edited UAN data, number of observations, and number of observations with SE.
Rate n n_with_se
Bin 1: 067-117 92 65
Bin 2: 118-167 76 73
Bin 3: 168-217 71 71
Bin 4: 218-268 54 51
A4.1.3 All combinations
The total number of combinations of values of the four factors fertilizer, apptime, place, and
rate was 6*4*4*4 = 384. Of these, the 70 listed in Table A 4.6 were represented in the data. Of
the 70, we included the subset of 50 treatment combinations for which there were at least 3
observations.
Table A 4.6 Fertilizer-apptime-place-rate combinations in the edited UAN data, number of observations, and number of observations with SE.
Combination Fertilizer apptime place rate n n_with_se
1 UAN+Nutrisphere AtPl broadcast Bin 1: 067-117 3 0
2 UAN+Nutrisphere AtPl broadcast Bin 2: 118-167 1 0
3 UAN+Nutrisphere AtPl broadcast Bin 4: 218-268 1 0
4 UAN+Nutrisphere AtPl surf band Bin 1: 067-117 2 0
5 UAN+Nutrisphere SD broadcast Bin 1: 067-117 1 0
6 UAN+Nutrisphere SD surf band Bin 1: 067-117 1 0
7 UAN AtPl bcast inc Bin 1: 067-117 9 9
8 UAN AtPl bcast inc Bin 2: 118-167 11 11
9 UAN AtPl bcast inc Bin 3: 168-217 5 5
10 UAN AtPl bcast inc Bin 4: 218-268 5 5
11 UAN AtPl broadcast Bin 1: 067-117 14 11
12 UAN AtPl broadcast Bin 2: 118-167 6 5
13 UAN AtPl broadcast Bin 3: 168-217 5 5
14 UAN AtPl broadcast Bin 4: 218-268 6 5
15 UAN AtPl surf band Bin 1: 067-117 2 0
16 UAN AtPl surf band Bin 3: 168-217 3 3
17 UAN AtPl surf band Bin 4: 218-268 1 1
Page 42 of 66
Combination Fertilizer apptime place rate n n_with_se
18 UAN AtPl/SD broadcast Bin 3: 168-217 4 4
19 UAN AtPl/SD broadcast Bin 4: 218-268 2 2
20 UAN PrePl bcast inc Bin 1: 067-117 2 2
21 UAN PrePl bcast inc Bin 2: 118-167 6 6
22 UAN PrePl broadcast Bin 1: 067-117 3 3
23 UAN PrePl broadcast Bin 3: 168-217 3 3
24 UAN SD bcast inc Bin 3: 168-217 1 1
25 UAN SD broadcast Bin 1: 067-117 2 0
26 UAN SD sub band Bin 1: 067-117 7 6
27 UAN SD sub band Bin 2: 118-167 5 5
28 UAN SD sub band Bin 3: 168-217 5 5
29 UAN SD sub band Bin 4: 218-268 5 5
30 UAN SD surf band Bin 1: 067-117 2 0
31 UAN SD surf band Bin 3: 168-217 3 3
32 UAN+cats AtPl bcast inc Bin 1: 067-117 5 5
33 UAN+cats AtPl bcast inc Bin 2: 118-167 5 5
34 UAN+cats AtPl bcast inc Bin 3: 168-217 5 5
35 UAN+cats AtPl bcast inc Bin 4: 218-268 5 5
36 UAN+cats AtPl broadcast Bin 1: 067-117 5 5
37 UAN+cats AtPl broadcast Bin 2: 118-167 5 5
38 UAN+cats AtPl broadcast Bin 3: 168-217 5 5
39 UAN+cats AtPl broadcast Bin 4: 218-268 5 5
40 UAN+Agrotain AtPl bcast inc Bin 1: 067-117 5 5
41 UAN+Agrotain AtPl bcast inc Bin 2: 118-167 5 5
42 UAN+Agrotain AtPl bcast inc Bin 3: 168-217 5 5
43 UAN+Agrotain AtPl bcast inc Bin 4: 218-268 5 5
44 UAN+Agrotain AtPl broadcast Bin 1: 067-117 5 5
45 UAN+Agrotain AtPl broadcast Bin 2: 118-167 5 5
46 UAN+Agrotain AtPl broadcast Bin 3: 168-217 5 5
47 UAN+Agrotain AtPl broadcast Bin 4: 218-268 5 5
48 UAN+Agrotain AtPl/SD broadcast Bin 2: 118-167 4 4
49 UAN+Agrotain AtPl/SD broadcast Bin 3: 168-217 4 4
Page 43 of 66
Combination Fertilizer apptime place rate n n_with_se
50 UAN+Agrotain AtPl/SD broadcast Bin 4: 218-268 2 2
51 UAN+Agrotain PrePl broadcast Bin 1: 067-117 3 3
52 UAN+Agrotain PrePl broadcast Bin 3: 168-217 3 3
53 UAN+AgrotainPlus AtPl bcast inc Bin 1: 067-117 5 5
54 UAN+AgrotainPlus AtPl bcast inc Bin 2: 118-167 5 5
55 UAN+AgrotainPlus AtPl bcast inc Bin 3: 168-217 5 5
56 UAN+AgrotainPlus AtPl bcast inc Bin 4: 218-268 5 5
57 UAN+AgrotainPlus AtPl broadcast Bin 1: 067-117 8 5
58 UAN+AgrotainPlus AtPl broadcast Bin 2: 118-167 6 5
59 UAN+AgrotainPlus AtPl broadcast Bin 3: 168-217 5 5
60 UAN+AgrotainPlus AtPl broadcast Bin 4: 218-268 6 5
61 UAN+AgrotainPlus AtPl surf band Bin 1: 067-117 2 0
62 UAN+AgrotainPlus AtPl surf band Bin 3: 168-217 3 3
63 UAN+AgrotainPlus AtPl surf band Bin 4: 218-268 1 1
64 UAN+AgrotainPlus SD broadcast Bin 1: 067-117 2 0
65 UAN+AgrotainPlus SD sub band Bin 1: 067-117 1 0
66 UAN+AgrotainPlus SD surf band Bin 1: 067-117 2 0
67 UAN+AgrotainPlus SD surf band Bin 3: 168-217 2 2
68 UAN+Instinct AtPl bcast inc Bin 2: 118-167 6 6
69 UAN+Instinct PrePl bcast inc Bin 1: 067-117 1 1
70 UAN+Instinct PrePl bcast inc Bin 2: 118-167 6 6
A4.1.4 UAN FARP3 data
Table A 4.7 lists the 50 treatment combinations in the 262-observation dataset we refer to as
UAN FARP3 (fertilizer, apptime, rate, and place with at least 3 observations). Rather than
further break down treatments according to rotation and tillage, we treated those variables as
random effects, due to a lack of sufficient observations for every possible combination. The
values of each of the variables in the UAN FARP 3 dataset were the same as the values in the
edited UAN dataset. All of the levels listed in
Table A 4.2, Table A 4.3, Table A 4.6, and
Table A 4.5 for each factor were still represented in the data.
Page 44 of 66
Table A 4.7 Fertilizer-apptime-place-rate combinations in the UAN FARP3 data, number of observations, and number of observations with SE.
Combination Fertilizer apptime place rate n n_with_se
1 UAN+Nutrisphere AtPl broadcast Bin 1: 067-117 3 0
2 UAN AtPl bcast inc Bin 1: 067-117 9 9
3 UAN AtPl bcast inc Bin 2: 118-167 11 11
4 UAN AtPl bcast inc Bin 3: 168-217 5 5
5 UAN AtPl bcast inc Bin 4: 218-268 5 5
6 UAN AtPl broadcast Bin 1: 067-117 14 11
7 UAN AtPl broadcast Bin 2: 118-167 6 5
8 UAN AtPl broadcast Bin 3: 168-217 5 5
9 UAN AtPl broadcast Bin 4: 218-268 6 5
10 UAN AtPl surf band Bin 3: 168-217 3 3
11 UAN AtPl/SD broadcast Bin 3: 168-217 4 4
12 UAN PrePl bcast inc Bin 2: 118-167 6 6
13 UAN PrePl broadcast Bin 1: 067-117 3 3
14 UAN PrePl broadcast Bin 3: 168-217 3 3
15 UAN SD sub band Bin 1: 067-117 7 6
16 UAN SD sub band Bin 2: 118-167 5 5
17 UAN SD sub band Bin 3: 168-217 5 5
18 UAN SD sub band Bin 4: 218-268 5 5
19 UAN SD surf band Bin 3: 168-217 3 3
20 UAN+cats AtPl bcast inc Bin 1: 067-117 5 5
21 UAN+cats AtPl bcast inc Bin 2: 118-167 5 5
22 UAN+cats AtPl bcast inc Bin 3: 168-217 5 5
23 UAN+cats AtPl bcast inc Bin 4: 218-268 5 5
24 UAN+cats AtPl broadcast Bin 1: 067-117 5 5
25 UAN+cats AtPl broadcast Bin 2: 118-167 5 5
26 UAN+cats AtPl broadcast Bin 3: 168-217 5 5
27 UAN+cats AtPl broadcast Bin 4: 218-268 5 5
28 UAN+Agrotain AtPl bcast inc Bin 1: 067-117 5 5
29 UAN+Agrotain AtPl bcast inc Bin 2: 118-167 5 5
Page 45 of 66
Combination Fertilizer apptime place rate n n_with_se
30 UAN+Agrotain AtPl bcast inc Bin 3: 168-217 5 5
31 UAN+Agrotain AtPl bcast inc Bin 4: 218-268 5 5
32 UAN+Agrotain AtPl broadcast Bin 1: 067-117 5 5
33 UAN+Agrotain AtPl broadcast Bin 2: 118-167 5 5
34 UAN+Agrotain AtPl broadcast Bin 3: 168-217 5 5
35 UAN+Agrotain AtPl broadcast Bin 4: 218-268 5 5
36 UAN+Agrotain AtPl/SD broadcast Bin 2: 118-167 4 4
37 UAN+Agrotain AtPl/SD broadcast Bin 3: 168-217 4 4
38 UAN+Agrotain PrePl broadcast Bin 1: 067-117 3 3
39 UAN+Agrotain PrePl broadcast Bin 3: 168-217 3 3
40 UAN+AgrotainPlus AtPl bcast inc Bin 1: 067-117 5 5
41 UAN+AgrotainPlus AtPl bcast inc Bin 2: 118-167 5 5
42 UAN+AgrotainPlus AtPl bcast inc Bin 3: 168-217 5 5
43 UAN+AgrotainPlus AtPl bcast inc Bin 4: 218-268 5 5
44 UAN+AgrotainPlus AtPl broadcast Bin 1: 067-117 8 5
45 UAN+AgrotainPlus AtPl broadcast Bin 2: 118-167 6 5
46 UAN+AgrotainPlus AtPl broadcast Bin 3: 168-217 5 5
47 UAN+AgrotainPlus AtPl broadcast Bin 4: 218-268 6 5
48 UAN+AgrotainPlus AtPl surf band Bin 3: 168-217 3 3
49 UAN+Instinct AtPl bcast inc Bin 2: 118-167 6 6
50 UAN+Instinct PrePl bcast inc Bin 2: 118-167 6 6
Total 262 248
A4.2 Model runs
Building a model designed to answer a question such as whether the effect of fertilizer on yield
was statistically significant requires making decisions and judgement calls regarding what to put
in the model and how to put it in. To examine the sensitivity of conclusions about the effect of
fertilizer to these decisions, we ran several different models.
A4.2.1 Estimability
With an observational dataset that had categorical variables with many levels, estimability of
treatment means must be considered. For example, for the 262-observation UAN dataset
described above, there were 6 levels of the factor fertilizer, and 4 levels of the factor apptime,
Page 46 of 66
but only 12 of the 24 combinations of values of fertilizer and apptime were in the
observational data. If all 24 combinations were available, it would be possible to consider in a
model the main effects of fertilizer and apptime and the interaction effect fertilizer*apptime.
When the factorial is not complete, however, it is possible to look at either the main effect of
fertilizer and apptime or the interaction effect, but not both. A model that contains the
interaction essentially treats each of the 12 combinations as a treatment.
A4.2.2 Categorical variable only model
To see if the conclusions about the effect of fertilizer on yield were sensitive to the inclusion of
the continuous variables, we began by fitting a categorical-variable-only UAN model, with
factors fertilizer, apptime, place, and rate—the categorical version of ratetot described above.
The model included a random effect of year nested within county and random effects of each of
rotate and till. Since the random effects of rotate and till were not significantly different from
zero, we tried a random effect of the interaction rotate*till. It was not significant either, so it
was eliminated from the model. The random effect of year within county was highly significant.
The p-values for fertilizer, apptime, place, and rate were, respectively, 0.99, 0.21, 0.07, and
<0.0001. Though at this point the effect of fertilizer was non-significant, we back-eliminated
other insignificant terms one at a time to obtain a model with place and rate with p-values 0.02
and <0.0001, respectively. Even though place was significant at the alpha = 0.05 level, the
LSMeans for each level of place were not significantly different from one another based on a
Tukey means comparison. There were significant differences in the LSMeans for some levels of
rate (Table A 4.8). These results were consistent with the expectation that yield increases and
then levels off as a function of N rate.
Table A 4.8 Tukey-Kramer grouping for rate least squares means (alpha=0.05). LSmeans with the same letter are not significantly different.
rate Estimated mean yield (Mg ha-1)
Bin 4: 218-268 11.7 A
Bin 3: 168-217 11.5 A
Bin 2: 118-167 10.8 B
Bin 1: 067-117 9.4 C
A4.2.3 Categorical and continuous variable models
We next put the factors fertilizer, apptime, and place into a model along with all of the
continuous variables, but with no interactions. Now N rate was being quantified as the
continuous variable erate. In this model, the p-values for fertilizer, apptime, place, and erate
were 0.97, 0.005, 0.005, and <0.0001. The addition of main effects of the continuous variables
Page 47 of 66
did not change the conclusion about the effect of fertilizer. Since there are times when a
variable has an effect on a response only by interacting with another variable, we added to the
above model interactions between each of the continuous variables and fertilizer. None of the
interaction terms had a p-value lower than 0.25. Including these interactions did not change the
conclusion that the effect of fertilizer is not significant.
We next fit a model with fertilizer, apptime, place, all of the continuous variables, and all 2-way
interactions between pairs of continuous variables that remained after the collinearity analysis.
The p-values for fertilizer, apptime, place, and erate in this model, which included both
categorical and continuous variables were 0.97, 0.0004, 0.0010, and <0.0001, respectively. We
did not include interactions between continuous and categorical variables—with the exception
of the exercise described in the previous paragraph—because there were not enough values of
each continuous variable within each level of each factor to allow reasonable estimation of
individual slopes. Figure A 4.1 illustrates this point. For at-planting, there was a wide range of
values of tempmax, but for AtPl/SD there were only two values of tempmax. To estimate the
slope of a line, it is best to have quite a few more than two values of the independent variable
represented to determine true underlying trends as opposed to over-fitting the data.
Page 48 of 66
Figure A 4.1 Scatterplot of yield vs. tempmax, with different colors/symbols for each value of apptime, and a different regression line for each value of apptime: at planting (AtPl; dark blue circles), at planting/sidedress (AtPl/SD; cyan crosses), preplant (PrePl; yellow triangles), and sidedress (SD; pink squares).
A4.2.4 Final model
From the categorical and continuous model with interactions, we back-eliminated all terms that
were not significant at the alpha = 0.01 level. We did not back-eliminate a main effect if there
was a significant interaction between it and another variable (Table A 4.9).
yld
4
5
6
7
8
9
10
11
12
13
14
15
16
17
tempmax
32 33 34 35 36 37 38 39 40
apptime AtPl AtPl/SD PrePl SD
Page 49 of 66
Table A 4.9 Type III tests of fixed effects for final UAN model.
Effect
Num
DF
Den
DF F Value Pr > F
Fertilizer 5 227.1 0.42 0.8376
apptime 3 234.2 4.95 0.0024
erate 1 236.6 134.84 <.0001
tempmax 1 24.18 12.38 0.0017
tempmin 1 35.22 6.73 0.0138
pH 1 177.7 17.02 <.0001
SOM 1 227.1 28.08 <.0001
tempmax_erate 1 239.4 7.69 0.0060
tempmin_erate 1 234.7 31.82 <.0001
ph_erate 1 232.2 22.42 <.0001
SOM_erate 1 236.2 10.04 0.0017
The LSMeans yield estimates and comparisons for the values of apptime (Table A 4.10) showed
that the mean yield for at-planting/sidedress was significantly higher than those for pre-
planting, at-planting, and sidedress, none of which were significantly different from the others.
Table A 4.10 Tukey-Kramer grouping for apptime least squares means (alpha=0.05). LSmeans with the same letter are not significantly different.
apptime Estimated mean yield (Mg ha-1)
AtPl/SD 11.1 A
PrePl 10.2 B
AtPl 9.4 B
SD 9.1 B
A4.3 Sensitivity to using weights
In meta-analyses in which the response variable is effect size (as opposed to treatment mean,
which we use), authors use weights such as 1/V, where V is the variance of the effect size
(Basche et al., 2014; Decock, 2014); r, the number of replicates supporting each treatment
(Linquist et al., 2013); 1/nstudy , where nstudy is the number of observations from the same study;
or some function of these (Decock, 2014; Pittelkow et al., 2015). How V is calculated depends
on the choice of effect size, which depends on the nature of the data and the objectives of the
researchers (Borenstein et al., 2009). Using 1/V gives more weight to effect sizes based on more
Page 50 of 66
precise treatment mean estimates. In our study, the response variable is not effect size, but
the treatment means of yield, and the variance of the treatment means is the square of the
standard error. Using 1/SE2 as a weighting factor is the analog to using 1/V. Using r as a
weighting factor gives more credit to treatment means based on more replicates, but it does
take into account relative precision for two different treatment means based on the same
number of replicates. The rationale for using 1/nstudy is to prevent any single study from having
an overwhelming influence on conclusions. That might be a good idea if every study covered a
single site and a single year, but different studies had different numbers of treatments, but that
is not the way studies are published. While one study might cover a single year at a single site
with several treatments, a second study might have the same treatments for a single year at
multiple sites, thus covering multiple sets of geographical, climatological, and soil conditions. A
third study might have the same treatments for multiple years at one site, covering multiple
sets of meteorological conditions, but with geographical, climatological, and soil conditions
fixed, and a fourth study might have the same treatments at multiple sites for multiple years,
covering multiple sets of geographical, climatological, meteorological, and soil conditions. Using
1/nstudy would give far too much weight to the observations from the first of the four
hypothetical studies, and far too little weight to the fourth. Of these choices of weights, using
1/V for effect sizes or 1/SE2 for treatment means are the only ones with a basis in probability
theory.
We examined to what extent conclusions about the effect of fertilizer on yield were sensitive to
the use of weights. From the UAN FARP3 data, we created a new 248-observation dataset
consisting of those treatment means for which SE was available (
Table A 4.7). We call this the UAN FARP3W data, where the W signifies that we are using this
dataset for a weighting exercise. We considered the weights w1 as 1/SE2 and w2 as r (Table A
4.11) because they have frequently been used in meta-analysis literature. We do not, however,
advocate use of r as a weight, since its use has no basis in probability theory. Table A 4.11
shows the quantiles for each weight in the UAN FARP3W data. The smallest value of 1/SE2 is 1,
while the largest is 400. Since the “weights” are applied in the fitting algorithm after the
squaring in least-squares occurs, and also after the squaring when maximum-likelihood or
pseudo-likelihood methods are used, the actual weight applied to the observation is √𝑤1 =
1/𝑆𝐸. Thus, using 1/SE2 results in weighting at least one observation 20 times more than at
least one other observation. The weight r does not produce vast differences in weights among
the observations since most field studies have similar numbers of replicates.
Page 51 of 66
Table A 4.11 Distribution of values of weights 1/SE2 and r in the UAN FARP3W dataset.
Variable Minimum 1st Pctl 10th Pctl Lower Quartile Median Upper Quartile 90th Pctl 99th Pctl Maximum
1 𝑆𝐸2⁄ 𝑟
1 3
1 3
2 3
3 4
5 4
12 4
23 4
83 5
400 5
Table A 4.12 gives the p-values of fertilizer for 5 different models. The first p-value is the same
as the one given in Section A4.2.4 for the model fit to the 262-observation UAN FARP3 data
with explanatory variables fertilizer, apptime, place, all of the continuous variables, all 2-way
interactions between pairs of continuous variables that made the cut after the collinearity
analysis, but no interactions between continuous and categorical variables. The reason for
using these terms was to go back to the last step prior to eliminating terms based on p-values,
because those p-values were calculated without weights.
The second p-value is for a model with all the same explanatory variables, but fit to the 248-
observation UAN FARP3W data, again with no weights. Comparing these two p-values shows
the difference in conclusions about the statistical significance of fertilizer that would result from
dropping the 14 observations. It is not surprising that there is very little difference, given that
the number of observations dropped is less than six percent of the total number of
observations.
The next two rows give the p-values resulting from fitting the same model using each of the
two weights. Comparing the p-value for the model fit to the UAN FARP3W data with no
weights, 0.9401, to that for the same data using w1, 0.9401, shows the difference in conclusions
about the statistical significance of fertilizer that results from weighting the observations with
w1. The statistical significance of fertilizer in the UAN model is not sensitive to the use of
weights.
Table A 4.12 P-values of fertilizer for UAN models that include the same explanatory variables, but different numbers of observations, or the same number of observations, but different weights.
Dataset n Weight Fertilizer p-value
UAN FARP3 262 none 0.9418
UAN FARP3W 248 none 0.9468
UAN FARP3W 248 w1 1/SE2 0.9401
UAN FARP3W 248 w2 r 0.9603
Page 52 of 66
We were able to perform this sensitivity analysis for the UAN and urea models because the
number of observations for which SE is available is close to the number when it is not (
Table A 4.7), and thus the coverage of treatment mean combinations is similar, and the
continuous predictor variable coverage is similar. We did not do such an analysis for AA
because the number of observations with SE is less than one third the total number of
observations in the AA data (
Table A 3.6). Thus we would have to build a completely different model to do a weight-
sensitivity exercise because the coverage of explanatory variables by the reduced dataset
would not support the model in Section A3.2.3. If we built an entirely new un-weighted model
based on the reduced dataset, we might see a difference in the statistical significance of
fertilizer due to eliminating 2/3 of the data. We could then apply weights and see if that
significance changed. We did not think it advisable to draw conclusions based on 1/3 of the
data when the weighting analysis suggested that it would not greatly change our conclusions.
We assume that if the statistical significance of the effect of fertilizer is not sensitive to
weighting for the UAN and urea models, it is unlikely that it would be for the AA model.
A4.4 Urea
A4.5 Selecting data
The third model focuses on urea as the nitrogen fertilizer source and includes several enhanced
efficiency fertilizers (Table A 4.13). Based on the availability of data, we classified Duration III,
ESN, and generic polymer coated fertilizer (PCF), and called the group “PCU.” We omitted the 4
observations where fertilizer = “urea/ESN,” since it was a mixture of ESN and urea, and there
were not enough observations to treat it as a separate fertilizer treatment. We omitted the
observation with fertilizer urea+nitrapyrin because the author did not specify which commercial
formulation of nitrapyrin was used, and we lumped the observations for urea+Instinct and
urea+N-Serve into one group which we called “urea+nitrapyrin.”
A4.5.1 Original Urea data
Table A 4.13 Source-inhibitor-fertilizer combinations in original Urea data, number of observations, and number of observations with SE.
Combination Source Inhibitor Fertilizer n n_with_se
1 urea NBPT urea+Agrotain 51 40
2 urea NBPT+DCD SuperU 78 74
Page 53 of 66
Combination Source Inhibitor Fertilizer n n_with_se
3 urea PCF Duration III 2 2
4 urea PCF ESN 160 118
5 urea PCF PCU 8 8
6 urea PCF urea/ESN 4 0
7 urea S.R Nutrisphere 17 14
8 urea nitrapyrin urea+Instinct 28 28
9 urea nitrapyrin urea+N-Serve 4 0
10 urea nitrapyrin urea+nitrapyrin 1 1
11 urea none urea 205 165
Total 558 450
A4.5.2 Edited Urea data
Table A 4.14 gives the results after the above-mentioned edits. In this 553-observation dataset,
the factor fertilizer had 6 levels, urea+Agrotain, SuperU, ESN, urea+Nutrisphere,
urea+nitrapyrin, and urea.
Table A 4.14 Values of the variable fertilizer in the edited Urea data, number of observations, and number of observations with SE.
Fertilizer n n_with_se
urea+Agrotain 51 40
SuperU 78 74
PCU 170 128
urea+Nutrisphere 17 14
urea+nitrapyrin 32 28
urea 205 165
Total 553
The factor apptime had 6 levels, at planting (AtPl), at planting/side dress (AtPl/SD), pre-planting
(PrePl), pre-planting/side dress (PrePl/SD), side dress (SD), and fall (Table A 4.15).
Table A 4.15 Values of the variable apptime in the edited urea data, number of observations, and number of observations with SE.
apptime n n_with_se
AtPl 319 257
Page 54 of 66
apptime n n_with_se
AtPl/SD 8 4
PrePl 103 81
PrePl/SD 3 3
SD 44 32
fall 76 72
The factor place had 4 levels: broadcast, broadcast incorporated (bcast inc), subsurface banded
(sub band), and surface banded (surf band) (Table A 4.16).
Table A 4.16 Values of the variable place in the edited urea data, number of observations, and number of observations with SE.
place n n_with_se
bcast inc 112 112
broadcast 134 114
sub band 23 21
surf band 24 13
The continuous variable ratetot had values from 56 to 276 kg ha-1. To allow a comparison of
coverage similar to that for the other 4R factors, we created the categorical variable rate to
divide ratetot into five evenly spaced bins from 56-96, 97-156, 157-216, 217-276, and 277-336
kg ha-1(Table A 4.17).
Table A 4.17 Values of the variable rate in the edited urea data, number of observations, and number of observations with SE.
rate n n_with_se
Bin 1: 56- 96 87 65
Bin 2: 97-156 208 182
Bin 3: 157-216 153 116
Bin 4: 217-276 79 70
Bin 5: 277-336 26 16
Total 553
A4.5.3 All combinations
The total number of combinations of the six levels of fertilizer, six levels of apptime, four levels
of place, and five levels of rate was 720. Of these combinations, the 127 that were represented
are listed in Table A 4.18.
Page 55 of 66
Table A 4.18 Fertilizer-apptime-place-rate combinations in the edited Urea data, number of observations, and number of observations with SE.
Combination Fertilizer apptime place rate n n_with_se
1 PCU AtPl bcast inc Bin 1: 56- 96 8 8
2 PCU AtPl bcast inc Bin 2: 97-156 15 11
3 PCU AtPl bcast inc Bin 3: 157-216 11 7
4 PCU AtPl bcast inc Bin 4: 217-276 5 5
5 PCU AtPl bcast inc Bin 5: 277-336 4 0
6 PCU AtPl broadcast Bin 1: 56- 96 11 8
7 PCU AtPl broadcast Bin 2: 97-156 11 9
8 PCU AtPl broadcast Bin 3: 157-216 10 8
9 PCU AtPl broadcast Bin 4: 217-276 7 5
10 PCU AtPl broadcast Bin 5: 277-336 1 0
11 PCU AtPl sub band Bin 2: 97-156 1 0
12 PCU AtPl sub band Bin 3: 157-216 1 1
13 PCU AtPl surf band Bin 3: 157-216 5 5
14 PCU AtPl surf band Bin 4: 217-276 6 6
15 PCU AtPl/SD broadcast Bin 2: 97-156 1 0
16 PCU AtPl/SD broadcast Bin 3: 157-216 1 0
17 PCU AtPl/SD broadcast Bin 4: 217-276 1 0
18 PCU AtPl/SD broadcast Bin 5: 277-336 1 0
19 PCU PrePl bcast inc Bin 1: 56- 96 1 1
20 PCU PrePl bcast inc Bin 2: 97-156 5 5
21 PCU PrePl bcast inc Bin 3: 157-216 9 3
22 PCU PrePl bcast inc Bin 4: 217-276 3 3
23 PCU PrePl bcast inc Bin 5: 277-336 5 5
24 PCU PrePl broadcast Bin 1: 56- 96 3 0
25 PCU PrePl broadcast Bin 2: 97-156 4 3
26 PCU PrePl broadcast Bin 3: 157-216 3 3
27 PCU PrePl sub band Bin 2: 97-156 1 0
28 PCU SD bcast inc Bin 3: 157-216 3 3
29 PCU SD broadcast Bin 1: 56- 96 1 0
30 PCU SD broadcast Bin 2: 97-156 3 2
Page 56 of 66
Combination Fertilizer apptime place rate n n_with_se
31 PCU SD sub band Bin 3: 157-216 1 1
32 PCU SD surf band Bin 3: 157-216 6 6
33 PCU fall bcast inc Bin 2: 97-156 4 4
34 PCU fall broadcast Bin 2: 97-156 6 5
35 PCU fall sub band Bin 1: 56- 96 2 2
36 PCU fall sub band Bin 2: 97-156 7 6
37 PCU fall surf band Bin 1: 56- 96 1 1
38 PCU fall surf band Bin 2: 97-156 2 2
39 SuperU AtPl bcast inc Bin 1: 56- 96 5 5
40 SuperU AtPl bcast inc Bin 2: 97-156 5 5
41 SuperU AtPl bcast inc Bin 3: 157-216 5 5
42 SuperU AtPl bcast inc Bin 4: 217-276 5 5
43 SuperU AtPl broadcast Bin 1: 56- 96 7 5
44 SuperU AtPl broadcast Bin 2: 97-156 5 5
45 SuperU AtPl broadcast Bin 3: 157-216 5 5
46 SuperU AtPl broadcast Bin 4: 217-276 5 5
47 SuperU AtPl surf band Bin 3: 157-216 3 3
48 SuperU AtPl surf band Bin 4: 217-276 3 3
49 SuperU AtPl/SD broadcast Bin 3: 157-216 1 1
50 SuperU PrePl bcast inc Bin 1: 56- 96 1 1
51 SuperU PrePl bcast inc Bin 2: 97-156 2 2
52 SuperU PrePl bcast inc Bin 3: 157-216 2 2
53 SuperU PrePl bcast inc Bin 4: 217-276 2 2
54 SuperU PrePl bcast inc Bin 5: 277-336 3 3
55 SuperU PrePl broadcast Bin 2: 97-156 2 2
56 SuperU PrePl broadcast Bin 3: 157-216 2 2
57 SuperU PrePl/SD broadcast Bin 2: 97-156 2 2
58 SuperU PrePl/SD broadcast Bin 3: 157-216 1 1
59 SuperU SD bcast inc Bin 3: 157-216 3 3
60 SuperU SD broadcast Bin 1: 56- 96 1 0
61 SuperU SD broadcast Bin 2: 97-156 3 2
62 SuperU SD surf band Bin 3: 157-216 5 5
Page 57 of 66
Combination Fertilizer apptime place rate n n_with_se
63 urea AtPl bcast inc Bin 1: 56- 96 8 8
64 urea AtPl bcast inc Bin 2: 97-156 25 21
65 urea AtPl bcast inc Bin 3: 157-216 11 7
66 urea AtPl bcast inc Bin 4: 217-276 4 4
67 urea AtPl bcast inc Bin 5: 277-336 4 0
68 urea AtPl broadcast Bin 1: 56- 96 9 6
69 urea AtPl broadcast Bin 2: 97-156 9 8
70 urea AtPl broadcast Bin 3: 157-216 10 8
71 urea AtPl broadcast Bin 4: 217-276 9 7
72 urea AtPl sub band Bin 2: 97-156 1 0
73 urea AtPl sub band Bin 3: 157-216 1 1
74 urea AtPl surf band Bin 3: 157-216 7 7
75 urea AtPl surf band Bin 4: 217-276 7 7
76 urea AtPl/SD broadcast Bin 3: 157-216 1 1
77 urea AtPl/SD surf band Bin 3: 157-216 2 2
78 urea PrePl bcast inc Bin 1: 56- 96 3 3
79 urea PrePl bcast inc Bin 2: 97-156 15 15
80 urea PrePl bcast inc Bin 3: 157-216 11 5
81 urea PrePl bcast inc Bin 4: 217-276 5 5
82 urea PrePl bcast inc Bin 5: 277-336 8 8
83 urea PrePl broadcast Bin 1: 56- 96 3 0
84 urea PrePl broadcast Bin 2: 97-156 1 0
85 urea PrePl sub band Bin 2: 97-156 1 0
86 urea SD bcast inc Bin 3: 157-216 3 3
87 urea SD broadcast Bin 1: 56- 96 1 0
88 urea SD broadcast Bin 2: 97-156 3 2
89 urea SD sub band Bin 3: 157-216 2 0
90 urea SD sub band Bin 4: 217-276 2 0
91 urea SD surf band Bin 3: 157-216 5 5
92 urea fall bcast inc Bin 2: 97-156 11 11
93 urea fall broadcast Bin 1: 56- 96 1 1
94 urea fall broadcast Bin 2: 97-156 8 7
Page 58 of 66
Combination Fertilizer apptime place rate n n_with_se
95 urea fall broadcast Bin 3: 157-216 1 1
96 urea fall broadcast Bin 4: 217-276 1 1
97 urea fall sub band Bin 1: 56- 96 2 2
98 urea fall sub band Bin 2: 97-156 7 6
99 urea fall surf band Bin 1: 56- 96 1 1
100 urea fall surf band Bin 2: 97-156 2 2
101 urea+Agrotain AtPl bcast inc Bin 1: 56- 96 5 5
102 urea+Agrotain AtPl bcast inc Bin 2: 97-156 5 5
103 urea+Agrotain AtPl bcast inc Bin 3: 157-216 9 5
104 urea+Agrotain AtPl bcast inc Bin 4: 217-276 5 5
105 urea+Agrotain AtPl broadcast Bin 1: 56- 96 8 5
106 urea+Agrotain AtPl broadcast Bin 2: 97-156 5 5
107 urea+Agrotain AtPl broadcast Bin 3: 157-216 6 5
108 urea+Agrotain AtPl broadcast Bin 4: 217-276 6 5
109 urea+Agrotain SD broadcast Bin 1: 56- 96 1 0
110 urea+Agrotain SD broadcast Bin 2: 97-156 1 0
111 urea+Nutrisphere AtPl bcast inc Bin 2: 97-156 2 2
112 urea+Nutrisphere AtPl broadcast Bin 1: 56- 96 2 1
113 urea+Nutrisphere AtPl broadcast Bin 2: 97-156 2 2
114 urea+Nutrisphere AtPl broadcast Bin 3: 157-216 2 1
115 urea+Nutrisphere AtPl broadcast Bin 4: 217-276 2 1
116 urea+Nutrisphere fall bcast inc Bin 2: 97-156 2 2
117 urea+Nutrisphere fall broadcast Bin 1: 56- 96 1 1
118 urea+Nutrisphere fall broadcast Bin 2: 97-156 2 2
119 urea+Nutrisphere fall broadcast Bin 3: 157-216 1 1
120 urea+Nutrisphere fall broadcast Bin 4: 217-276 1 1
121 urea+nitrapyrin AtPl bcast inc Bin 2: 97-156 5 5
122 urea+nitrapyrin AtPl bcast inc Bin 3: 157-216 4 0
123 urea+nitrapyrin AtPl broadcast Bin 2: 97-156 2 2
124 urea+nitrapyrin PrePl bcast inc Bin 1: 56- 96 1 1
125 urea+nitrapyrin PrePl bcast inc Bin 2: 97-156 7 7
Page 59 of 66
Combination Fertilizer apptime place rate n n_with_se
126 urea+nitrapyrin fall bcast inc Bin 2: 97-156 7 7
127 urea+nitrapyrin fall broadcast Bin 2: 97-156 6 6
A4.5.4 Urea FARP3 data
Of those 127 combinations, we chose the 74 for which there were at least 3 observations. We
call this the Urea FARP3 dataset. There are a total of 479 observations in these combinations
(Table A 4.19).
Table A 4.19 Fertilizer-apptime-place-rate combinations in the edited urea FARP3 data, number of observations, and number of observations with SE.
Combination Fertilizer apptime place rate n n_with_se
1 PCU AtPl bcast inc Bin 1: 56- 96 8 8
2 PCU AtPl bcast inc Bin 2: 97-156 15 11
3 PCU AtPl bcast inc Bin 3: 157-216 11 7
4 PCU AtPl bcast inc Bin 4: 217-276 5 5
5 PCU AtPl bcast inc Bin 5: 277-336 4 0
6 PCU AtPl broadcast Bin 1: 56- 96 11 8
7 PCU AtPl broadcast Bin 2: 97-156 11 9
8 PCU AtPl broadcast Bin 3: 157-216 10 8
9 PCU AtPl broadcast Bin 4: 217-276 7 5
10 PCU AtPl surf band Bin 3: 157-216 5 5
11 PCU AtPl surf band Bin 4: 217-276 6 6
12 PCU PrePl bcast inc Bin 2: 97-156 5 5
13 PCU PrePl bcast inc Bin 3: 157-216 9 3
14 PCU PrePl bcast inc Bin 4: 217-276 3 3
15 PCU PrePl bcast inc Bin 5: 277-336 5 5
16 PCU PrePl broadcast Bin 1: 56- 96 3 0
17 PCU PrePl broadcast Bin 2: 97-156 4 3
18 PCU PrePl broadcast Bin 3: 157-216 3 3
19 PCU SD bcast inc Bin 3: 157-216 3 3
20 PCU SD broadcast Bin 2: 97-156 3 2
21 PCU SD surf band Bin 3: 157-216 6 6
22 PCU fall bcast inc Bin 2: 97-156 4 4
Page 60 of 66
Combination Fertilizer apptime place rate n n_with_se
23 PCU fall broadcast Bin 2: 97-156 6 5
24 PCU fall sub band Bin 2: 97-156 7 6
25 SuperU AtPl bcast inc Bin 1: 56- 96 5 5
26 SuperU AtPl bcast inc Bin 2: 97-156 5 5
27 SuperU AtPl bcast inc Bin 3: 157-216 5 5
28 SuperU AtPl bcast inc Bin 4: 217-276 5 5
29 SuperU AtPl broadcast Bin 1: 56- 96 7 5
30 SuperU AtPl broadcast Bin 2: 97-156 5 5
31 SuperU AtPl broadcast Bin 3: 157-216 5 5
32 SuperU AtPl broadcast Bin 4: 217-276 5 5
33 SuperU AtPl surf band Bin 3: 157-216 3 3
34 SuperU AtPl surf band Bin 4: 217-276 3 3
35 SuperU PrePl bcast inc Bin 5: 277-336 3 3
36 SuperU SD bcast inc Bin 3: 157-216 3 3
37 SuperU SD broadcast Bin 2: 97-156 3 2
38 SuperU SD surf band Bin 3: 157-216 5 5
39 urea AtPl bcast inc Bin 1: 56- 96 8 8
40 urea AtPl bcast inc Bin 2: 97-156 25 21
41 urea AtPl bcast inc Bin 3: 157-216 11 7
42 urea AtPl bcast inc Bin 4: 217-276 4 4
43 urea AtPl bcast inc Bin 5: 277-336 4 0
44 urea AtPl broadcast Bin 1: 56- 96 9 6
45 urea AtPl broadcast Bin 2: 97-156 9 8
46 urea AtPl broadcast Bin 3: 157-216 10 8
47 urea AtPl broadcast Bin 4: 217-276 9 7
48 urea AtPl surf band Bin 3: 157-216 7 7
49 urea AtPl surf band Bin 4: 217-276 7 7
50 urea PrePl bcast inc Bin 1: 56- 96 3 3
51 urea PrePl bcast inc Bin 2: 97-156 15 15
52 urea PrePl bcast inc Bin 3: 157-216 11 5
53 urea PrePl bcast inc Bin 4: 217-276 5 5
54 urea PrePl bcast inc Bin 5: 277-336 8 8
Page 61 of 66
Combination Fertilizer apptime place rate n n_with_se
55 urea PrePl broadcast Bin 1: 56- 96 3 0
56 urea SD bcast inc Bin 3: 157-216 3 3
57 urea SD broadcast Bin 2: 97-156 3 2
58 urea SD surf band Bin 3: 157-216 5 5
59 urea fall bcast inc Bin 2: 97-156 11 11
60 urea fall broadcast Bin 2: 97-156 8 7
61 urea fall sub band Bin 2: 97-156 7 6
62 urea+Agrotain AtPl bcast inc Bin 1: 56- 96 5 5
63 urea+Agrotain AtPl bcast inc Bin 2: 97-156 5 5
64 urea+Agrotain AtPl bcast inc Bin 3: 157-216 9 5
65 urea+Agrotain AtPl bcast inc Bin 4: 217-276 5 5
66 urea+Agrotain AtPl broadcast Bin 1: 56- 96 8 5
67 urea+Agrotain AtPl broadcast Bin 2: 97-156 5 5
68 urea+Agrotain AtPl broadcast Bin 3: 157-216 6 5
69 urea+Agrotain AtPl broadcast Bin 4: 217-276 6 5
70 urea+nitrapyrin AtPl bcast inc Bin 2: 97-156 5 5
71 urea+nitrapyrin AtPl bcast inc Bin 3: 157-216 4 0
72 urea+nitrapyrin PrePl bcast inc Bin 2: 97-156 7 7
73 urea+nitrapyrin fall bcast inc Bin 2: 97-156 7 7
74 urea+nitrapyrin fall broadcast Bin 2: 97-156 6 6
Total 479 397
A4.6 Model runs
A4.6.1 Categorical variable only model
As with the UAN model, to see if the conclusions about the effect of fertilizer on yield were
sensitive to the inclusion of the continuous variables, we began by fitting a categorical-variable-
only urea model, with factors fertilizer, apptime, place, and rate—the categorical version of
ratetot described above. The model included a random effect of year nested within county and
random effects of each of rotate and till. Since the random effects of rotate and till were not
significantly different from zero, we tried a random effect of the interaction rotate*till. It was
not significant and eliminated from the model. The random effect of year within county was
highly significant. The p-values for fertilizer, apptime, place, and rate are shown in Table A 4.20.
Page 62 of 66
Table A 4.20 Type III tests of fixed effects for categorical variable only urea model.
Effect Num DF Den DF F Value Pr > F
Fertilizer 4 424.3 1.59 0.1747
apptime 3 447.5 4.09 0.0069
place 3 445.7 2.17 0.0908
rate 4 438.1 28.61 <.0001
The LSMeans tables for apptime and rate from this model show that mean yield for sidedress
application was significantly higher than that for pre-plant and fall applications, but was not
significantly different from at planting applications (Table A 4.21). There were no significant
differences in mean yield among at planting, pre-planting, or fall applications.
Table A 4.21 Tukey-Kramer grouping for apptime least squares means (Alpha=0.05). LS-means with the same letter are not significantly different.
apptime Estimated mean yield (Mg ha-1)
SD 12.8 A
AtPl 11.0 B A
PrePl 10.6 B
fall 10.5 B
Mean yield for Bin 5, representing the highest N rates, was significantly higher than those for
Bins 1-4. Mean yield for Bin 4 was significantly higher than for Bins 2 and 1, but was not
significantly different from Bin 3. Mean yields for Bins 2 and 3 were significantly higher than for
Bin 1, but were not significantly different from one another.
Table A 4.22 Tukey-Kramer grouping for rate least squares means (Alpha=0.05). LS-means with the same letter are not significantly different.
rate Estimated mean yield (Mg ha-1)
Bin 5: 277-336 13.1 A
Bin 4: 217-276 11.6 B
Bin 3: 157-216 11.3 C B
Bin 2: 97-156 10.7 C
Bin 1: 56- 96 9.4 D
Page 63 of 66
A4.6.2 Categorical and continuous
We next put the factors fertilizer, apptime, and place into a model along with all of the
continuous variables, but with no interactions. In this model, we used N rate as the continuous
variable erate (where rate was modified to an exponential functional form). In this model, the
p-values for fertilizer, apptime, place, and erate were 0.058, <0.0001, 0.042, and <0.0001.
At this stage for UAN, we considered a model with fertilizer, apptime, place, erate, all other
continuous variables, and interactions of fertilizer with all of the continuous variables. This was
not possible for urea, however, because the only way to include interactions between fertilizer
and the continuous variables would have been to omit the main effects of the continuous
variables from the model. There were not enough unique values of each continuous variable for
each level of the variable fertilizer, which, as was the case for UAN, was also the reason
interactions between the other categorical variables and the continuous variables were not
included.
We next fit a model with fertilizer, apptime, place, all of the continuous variables, and all 2-way
interactions between pairs of continuous variables that remained after the collinearity analysis.
As noted above, this model did not include interactions between categorical and continuous
variables. The p-values for fertilizer, apptime, place, and erate in the model were 0.029,
<0.0001, 0.0093, and <0.0001, respectively.
A4.6.3 Final model
Finally, we back-eliminated from this model all terms that were not significant at the alpha =
0.01 level. We performed this elimination in blocks, eliminating a few variables at a time. A
main effect would not be eliminated, however, if an interaction of that effect with another
variable remained in the model. After performing the back-eliminations, we examined residual
plots to assess model fit and to determine whether any observations exerted undue influence
on results. We found a distinct downward-trend-pattern in the plot of residuals vs. predicted
values that indicated a problem with the fit. Our investigation showed that outlying values of
sand and clay caused variables formed as interactions between sand and clay and other
continuous variables to appear to have statistical significance, but inclusion of these terms in
the model created the pattern mentioned. We added and subtracted terms from the model and
examined residual plots and statistical significance of the remaining terms. We also
reconsidered the random effect rotate*till, which is significant based on the 95% confidence
bounds. We arrived at a final model that included main effects of all of the continuous
variables, all of which were significant at the alpha = 0.01 level, except for tempmax, which was
very close to being significant at this level (Table A 4.23). The final model includes no
interactions, but does include apptime, which was highly significant, the random effect of year
nested within county, and the random effect rotate*till. As with the other sources, we left
Page 64 of 66
fertilizer in the final model. Fertilizer was not significant at the alpha = 0.05 level, and, as the
fertilizer LSMeans table demonstrates (Table A 4.24), there were no significant differences in
the mean yield for the different urea-based fertilizers.
Table A 4.23 Type III tests of fixed effects for final urea model.
Effect Num DF Den DF F Value Pr > F
Fertilizer 4 385.8 2.36 0.0525
apptime 3 430.8 6.15 0.0004
tempmax 1 34.38 7.39 0.0102
tempmin 1 43 7.69 0.0082
pH 1 291.6 7.84 0.0055
SOM 1 425.6 16.46 <.0001
totalwater 1 167.8 47.60 <.0001
erate 1 388.2 148.04 <.0001
Table A 4.24 Tukey-Kramer grouping for Fertilizer least squares means (alpha=0.05). LSmeans with the same letter are not significantly different.
Fertilizer Estimated mean yield (Mg ha-1)
SuperU 10.3 A
PCU 10.2 A
urea+Agrotain 10.2 A
urea+nitrapyrin 9.8 A
urea 9.8 A
In the urea model, mean yield for sidedress applications was significantly higher than for pre-
plant and fall applications, but was not significantly different from that for at planting
applications (Table A 4.25). Mean yield for at planting applications was significantly higher than
that for pre planting applications, but was not significantly different from fall applications.
Mean yield for fall and pre planting applications were not significantly different. Interpretation
of these results should be with care as the dataset was developed to model enhanced efficiency
fertilizers, not to quantify effects of application timing on yield.
Page 65 of 66
Table A 4.25 Tukey-Kramer grouping for apptime least squares means (alpha=0.05). LSmeans with the same letter are not significantly different.
apptime Estimated mean yield (Mg ha-1)
SD 11.9 A
AtPl 10.0 B A
fall 9.4 B C
PrePl 9.0 C
A4.7 Sensitivity to weights
Just as we did for the UAN model, we performed an analysis to test the sensitivity of the effect
of fertilizer on yield based to the use of weights. From the 479-observation Urea FARP3 data
set, we created a new 397-observation dataset consisting of those treatment means for which
SE was available. We call this the urea FARP3W data, where the W signifies that we are using
this dataset for a weighting exercise.
We again considered the weights: 1/SE2 and r. Quantiles for each weight in the Urea FARP3W
data show that the distribution of values of w1 are more disparate for the urea data than for
the UAN data (Table A 4.26). The maximum value is still 400, but the minimum value is 0.2, so
that the observation given weight 400 will be given √2000 ≈ 45 times the weight of the
observation given weight 0.2. The reason that r, the number of replicates, takes a maximum
value of 12 is that there are some treatment means that were reported by the authors as
averages over years. Thus those means are based on the sum of the replicates over those years.
Table A 4.26 Distribution of values of weights 1/SE2 and r in the urea FARP3W dataset.
Variable Minimum
1st
Pctl
10th
Pctl
Lower
Quartile Median
Upper
Quartile
90th
Pctl
99th
Pctl Maximum
1/SE2
r 0.2
3.0
0.3
3.0
1.5
3.0
2.8
4.0
5.2
4.0
12.8
4.0
27.7
4.0
277.8
12.0
400.0
12.0
The p-values of fertilizer for five different models are shown in Table A 4.27. The first p-value,
0.0572 is for a model including fertilizer, apptime, place, and main effects of all continuous
variables. It differs from the urea final model only in that we added place back to this model,
since place had been eliminated based on non-weighted p-values. The second p-value, 0.0738,
is for a model with all the same explanatory variables, but fit to the 397-observation urea
FARP3W data, again with no weights. Comparing the first two p-values shows the difference in
Page 66 of 66
conclusions that would result from dropping the 82 observations: the effect of fertilizer was
not significant at the alpha=0.05 level before or after removing the observations.
The next two rows give the p-values resulting from fitting the same model using the two
weights. We can see that for a model fit to the urea FARP3W data, going from using no weight
to using 1/SE2 changed the p-value from 0.0738 to 0.0465. The latter p-value would indicate
statistical significance at the alpha = 0.05 level. Fertilizer would probably be eliminated if we
took this model and back-eliminated until all terms were significant at the alpha = 0.01 level, as
we did with the urea final model. The reason we used the alpha = 0.01 level was that we are
performing many tests in all of these models for the sensitivity analyses and backward
elimination steps – far more hypothesis tests than one would perform when analyzing the
results of a designed experiment. Reducing the statistical significance level reduces the number
of tests that would be incorrect. Using the alpha = 0.01 level, the significance of fertilizer was
not sensitive to the use of weights.
Table A 4.27 P-values of fertilizer for UAN models that include the same explanatory variables, but different numbers of observations, or the same number of observations, but different weights.
Dataset Weight Fertilizer p-value
UREA FARP3 none 0.0572
UREA FARP3W none 0.0738
UREA FARP3W 1/SE2 0.0465
UREA FARP3W r 0.0841